Valerie M. Hudson - Artificial Intelligence and International Politics-Westview Press (1991)

Artificial Intelligence
and International Politics

and International Politics
edited by
Valerie M. Hudson
First published 1991 by Westview Press, Inc.
Published 2018 by Routledge

52 Vanderbilt Avenue, New York, NY 10017
2 Park Square, Milton Park, Abingdon, Oxon OX14 4RN
Routledge is an imprint of the Taylor & Francis Group, an informa business
Copyright © 1991 Taylor & Francis
All rights reserved. No part of this book may be reprinted or reproduced or utilised in any form
or by any electronic, mechanical, or other means, now known or hereafter invented, including
photocopying and recording, or in any information storage or retrieval system, without
permission in writing from the publishers.
Notice:
Product or corporate names may be trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.
Library of Congress Cataloging-in-Publication Data

Artificial intelligence and international politics I edited by Valerie
M. Hudson.
p. em.
Includes bibliographical references and index.
ISBN 0-8133-0937-9
1. International relations-Research. 2. International relations-
Data processing. 3. Artificial intelligence. I. Hudson, Valerie
M.
JX1291.A73 1991
327'.072-dc20 90-43085
CIP
ISBN 13: 978-0-367-00371-5 (hbk)

CONTENTS
About the Contributors vii
Introduction, Valerie M. Hudson 1
1 Artificial Intelligence and International Relations:

An Overview, Philip A. Schrodt 9
PART ONE
CONCEPTUAL ISSUES AND PRACTICAL CONCERNS
2 Artificial Intelligence and Intuitive Foreign Policy

Decision-Makers Viewed as Limited Information
Processors: Some Conceptual Issues and
Practical Concerns for the Future, He/en E. Purkitt 35
3 Steps Toward Artificial Intelligence: Rule-Based, Case-Based,

and Explanation-Based Models of Politics, Dwain Mefford 56
4 Text Modeling for International Politics:

A Tourist's Guide to RELATOS, Hayward R. Alker, Jr.,
Gavan Duffy, Roger Hurwitz, and John C. Mallery 97
5 Reasoning and Intelligibility, James P. Bennett and

Stuart J. Thorson 127
6 The Computational Modeling of Strategic Time,

Howard Tamashiro 149
PART TWO
AI/IR RESEARCH: INTERNATIONAL EVENTS AND
FOREIGN POLICY DECISION MAKING
7 Pattern Recognition of International Event Sequences:

A Machine Learning Approach, Philip A. Schrodt 169
V
vi Contents
8 Scripting International Power Dramas: A Model

of Situational Predisposition, Valerie M. Hudson 194
9 UNCLESAM: The Application of a Rule-Based Model

of U.S. Foreign Policy Making, Brian L. Job and
Douglas Johnson 221
10 Modeling Foreign Policy Decision Making as

Knowledge-Based Reasoning, Donald A. Sylvan,
Ashok Goe/, and B. Chandrasekaran 245
11 Decision Making and Development: A "Glass Box"

Approach to Representation, Margee M. Ensign and
Warren R. PhjJJips 274
PART THREE
AI/IR RESEARCH: THE DISCOURSE OF FOREIGN POLICY
12 The Expertise of the Senate Foreign Relations Committee,

G. R. Boynton 291
13 Reproduction of Perception and Decision in the

Early Cold War, Sanjoy Banerjee 310
14 Theoretical Categories and Data Construction in

Computational Models of Foreign Policy, David J. Sylvan,
Stephen J. Majeski, and Jennifer L. Mj/Jiken 327
15 Semantic Content Analysis: A New Methodology for the

RELATUS Natural Language Environment, John C. Mallery 347
16 Time Space: Representing Historical Time for

Efficient Event Retrieval, Gavan Duffy 386
About the Book and Editor 407

Index 409
CONTRIBUTORS
Hayward R. Alker, Jr., is a Professor of Political Science at MIT.

Sanjoy Banerjee is an Assistant Professor in the International Relations
Program at San Francisco State University.
James P. Bennett is an Associate Professor of Political Science at Syracuse
University.
G. R. Boynton is a Professor of Political Science at the University of Iowa.
B. Chandrasekaran is a Professor of Computer and Information Sciences
at the Ohio State University.
Gavan Duffy is Assistant Professor of Political Science at Syracuse University.
Margee M. Ensign is an Assistant Professor of Political Science at Columbia
University and the coordinator of the International Political Economy Program
in the School of International and Public Affairs.
Ashok Goel is an Assistant Professor of Computer and Information Sciences
at Georgia Technical Institute.
Valerie M. Hudson is an Assistant Professor of Political Science and the
Director of Graduate Studies at the David M. Kennedy Center for International
and Area Studies at Brigham Young University.
Roger Hurwitz is a doctoral candidate in Political Science at MIT.
Brian L. Job is a Professor of Political Science at the University of British
Columbia.
Douglas Johnson is a free-lance computer programmer based in Minnesota.
Stephen J. Majeski is an Assistant Professor of Political Science at the
University of Washington.
John C. Mallery is a doctoral candidate in Political Science and Electrical
Engineering and Computer Science at MIT.
Dwain Mefford is an Assistant Professor of Political Science at the Ohio
State University.
vii
viii Contributors
Jennifer L. Milliken is a doctoral candidate in Political S.cience at the

University of Minnesota.
Warren R. Phillips is a Professor of International Relations in the Department
of Government and Politics at the University of Maryland.
Helen E. Purkitt is an Associate Professor of Political Science at the U.S.
Naval Academy.
Philip A. Schrodt is an Associate Professor of Political Science at the
University of Kansas.
David J. Sylvan is an Associate Professor of Political Science at the University
of Minnesota.
Donald A. Sylvan is an Associate Professor of Political Science at the Ohio
State University.
Howard Tamashiro is an Assistant Professor of Political Science at Allegheny
College.
Stuart J. Thorson is a Professor in, and Chairman of, the Department of
Political Science at Syracuse University.
Introduction
Valerie M. Hudson
The purpose of this volume is to provide a general yet relatively com-

prehensive overview of the study of international relations using a com-
putational modeling approach (herein abbreviated Al/IR). 1 Such an approach
incorporates sophisticated, formal models of human reasoning into expla-
nations of human behavior. In the international relations context, this means
to offer explanations of international behavior in terms of the reasoning,
inference, and language use of relevant humans-whether these be national
leaders, foreign policy decision-making groups, or even scholars of inter-
national relations themselves.
At this stage in its development, separating the broader theoretical
objectives of AI/IR from the more personal motivations of those doing the
research is difficult. This is in part because only a relatively small number
of scholars are self-consciously working in the AI/IR subfield, and their
research aims have thus far served to define AI/IR for others. At some
point in the growth of this subfield a separation will be easier to make.
In the meantime, it is instructive to trace the general route that led most
AI/IR researchers to the point this edited volume represents. By and large,
most of the contributors to the volume began their careers developing and
applying mathematical and/or statistical models to "hard" (read "quantifi-
able") data in JR. To use the jargon, they would be seen as QJPers-QIP
standing for "quantitative international politics."2 Hayward Alker at one point
in his career searched for statistical patterns in Onited Nations voting
behavior. Philip Schrodt elaborated mathematical models of arms races.
Donald Sylvan used autoregressive integrated moving averages to explain
the dynamics of arms transfers between nations. Stuart Thorson once wrote
of visualizing the process of foreign policy adaptation as a chreod. And so
on, with few exceptions.
Dissatisfaction crept in, however. The central theme of articles expressing
dissatisfaction (see, for example, Alker and Chrlstensen, 1972; Schrodt,
1984) was the inadequate expressibility inherent in the way they had been
1
2 Valerie M. Hudson
Figure 1: Weinberg 's Classification of Subjects (From Weinberg, 1975, p. 18)
Ill. Orgaruud Complexity

(systems)
analytical treatment
COMPLEXITY statistical treatment F''>'>>'
',',. .,. .,. .,. .,. .,l
studying international relations. This problem had numerous facets: inadequate

expressibility of human reasoning and intentionality, of the richness of
qualitative IR data, of the nonrandom but untidy patterns found in human
behavior. Above all, what was most troublesome about the methods they
employed was that they seemed incapable of offering an explanation of
human behavior in IR that was fundamentally satisfying. The remark,
attributed to Rudolph Rummel, to the effect of " if your r 2 is 1.0, then you
have explained the phenomenon" did not sit well with some scholars who
were, in fact, busy calculating r 2 s themselves.
This problem is not unique to IR: lt is endemic to the social sciences.
Philosophers of social science have amplified this dissatisfaction for decades
now (see Taylor, 1985; Ham§ and Secord, 1973; Harre, 1984). When you
bring the human being or collectivities of humans back in as the focus of
research,3 as these and other philosophers would have us do, the richness
of human experience and human reasoning overpowers our methods. This
is ironic because for quite some time now quantitative methods have been
viewed as the most powerful methods available to the social scientist.
Gerald Weinberg's conceptualization of the problem remains a very useful
one (see Figure 1). He breaks down realms of scientific inquiry into three
regions according to the amount of randomness and complexity found therein
(see Weinberg, 1975).
Introduction 3
In Region I, which Weinberg calls the realm of "organized simplicity," or

of "mechanisms," there is enough structure and a small enough population
to allow analytic or mathematical methods to accurately and completely
explain and predict phenomena. In Region 11, the region of "unorganized
complexity," there is such a lack of structure (or, alternatively, presence of
randomness), and such large N sizes that statistical methods are the
appropriate tool of inquiry. However, in Region Ill, that of "organized
complexity," lie all the realms too complex for analytical treatment and too
structured for statistical treatment. The realm of human experience, decision
making, and behavior is most profitably seen as such a "medium-number
system." lt is quite possible to apply the methodology of Regions I and 11
to understand humans and their behavior, but Weinberg likens this to using
a band saw on one's fingernails:
[A] band saw is [not] responsible for the consequences of its being used to
trim fingernails. If fingernails need cutting, and the band saw is the only
available cutting tool, then the results are more or less predictable. A band
saw is a most useful tool, but not for certain jobs (Weinberg, 1975, 20).
The search for more refined tools of greater expressive power led some
to paths quite foreign: computer science, information processing, social
psychology, linguistic analysis, hermeneutics, and even a return to the more
traditional case study approach. Such journeys can be traced in the con-
tributions to this volume.
Two observations merit making at this point. The first is that, with some
exceptions, 4 there is still commitment to what we might call certain "scientific
desiderata" among those who do AI/IR. One reason that the transition from
QJP to AI/IR was not painful was that the techniques of the latter enabled
researchers to keep the rigor, explicitness, and overt evaluation criteria of
the former. The attractiveness to some of an approach capable of expressing
the intricacies of human reasoning and experience while at the same time
affording some vision of scientific progression should not be understated.
The second observation is that the odyssey made by many who have
contributed to this volume is, in a fundamental sense, a very familiar one.
lt is common to all forms of human inquiry to seek means of expressing
puzzles or problems that could not previously be expressed. At some point,
scientists become inventors: They invent the world they personally wish to
explore. I see the pieces in this volume as the work of inventors, adapting
existing tools and constructing new ones in order to see what is to them
more interesting than what they could see with the tools at hand. Robert
Root-Bernstein, a biochemist and historian of science, could easily have
written of AI/IR when he stated,
4 Valerie M. Hudson
It is not science's ability to reach solutions that needs to be accounted for,

but its ability to define solvable problems. . . . [l)nduction, deduction, and
abduction will not suffice to solve the range of problems scientists address.
Scientific "tools of thought" are more diverse than this and have been developed
not only to reason and to test, but to invent. These tools of invention include
(and are probably not limited to): abstracting, modeling, analogizing, pattern
forming and pattern recognition, aesthetics, visual thinking, playacting, and
manipulative skill. I suggest that scientists might better educate their successors
if they included these tools of thought in teaching their science (Root-Bernstein,
1989a, 486).
As with any odyssey, there are pitfalls, dead ends, wrong turns. AI/IR
represents no magical shortcut to the explanation of international relations.
Indeed, Charles Taylor's taunt of several years ago should still bother those
who choose this journey:
Theories [such as those found in artificial intelligence] lead to very bad science:
either they end up in wordy elaborations of the obvious, or they fail altogether
to address the interesting questions, or their practitioners end up squandering
their talents and ingenuity in the attempt to show that they can after all
recapture the insights of ordinary life in their manifestly reductive explanatory
languages (Taylor, 1985, 1).
What is so very ironic is that AI/IR does wish to recapture the insights
of ordinary life, to wit, how humans reason about international relations.
Furthermore, use of a computer does consign one at this point to a
"manifestly reductive language." What remains to be seen, pace Taylor, is
the value of reintroducing to IR the study of how policy-makers and scholars
reason about IR and explain IR to themselves. I suspect that value to be
quite high. As Mefford argues in this volume, to provide satisfying explanation
in international relations, we must strive for "strong" theory. Such theory,
in a domain devoted to the study of intentional behavior produced by human
reasoning as is international relations, would produce models that capture
this process as realistically as possible. AI/IR models go a long way toward
realizing that objective and, by so doing, progress what we are capable of
offering as explanation in international relations.
In my opinion, AI/IR bridges the gap between QJP and traditional IR by
merging the strengths of the former with the strengths of the latter. (See
Hudson, 1987.) Now one can speak of modeling, in a rigorous and explicit
manner, the richness of IR as a human activity, with all that that adjective
implies. To begin to accomplish that, even at first by means of a reductionist
technology, could be revolutionary in the study of JR. lt may well be the
catalyst for a whole new class of theory-building efforts in international
Introduction 5
relations. At the very least, the efforts of those who head in this direction
deserve serious scrutiny and reflection-hence this volume.
Philip Schrodt opens the volume by tracing the evolution of AI/IR by
reference to the evolution of AI in general. The remainder of the volume
is divided into two major sections: Conceptual Issues and Practical Concerns,
and AI/IR Research.
The first section is designed to acquaint the reader with general theoretical
and applied issues that are of interest and relevance to the subfield. Helen
Purkitt's chapter reinforces the shift to a computational modeling approach
in IR by surveying recent empirical findings in social psychology (and related
fields), cataloguing what is known about the idiosyncracies of human reasoning
and perception. Dwain Mefford's piece, which follows, takes the reader on
an informative reconnaissance of AI approaches: rule-based systems, case-
based systems, and explanation-based learning systems. Mefford compares
their strengths and weaknesses by constructing a model of political reasoning
about coups d'etat using each of the three approaches. Hayward Alker and
his coauthors then introduce the study of "natural language processing"
(NLP), which can be seen as distinct from much of the work in the volume
in terms of its objectives. NLP work endeavors to construct computer-aided
tools for the systematic processing of what some consider to be the heart
of politics-textual accounts of political phenomena. James Bennett and
Stuart Thorson also explore how natural language accounts of political
concepts-specifically, deterrence-can be formalized using computer lan-
guages. This formalization allows for a type of analysis impossible using
other methods. Rounding out this first section is Howard Tamashiro's chapter,
which illustrates how extremely important strategic concepts (here, time)
are often neglected in the absence of techniques capable of capturing their
nuances, which techniques can be found in computational modeling.
The second section of the volume showcases actual research (completed
or ongoing) produced by those working in the AI/IR subfield. This section
is subdivided into two parts to distinguish different broad categories of AI/
IR research. It should be noted that these subdivisions are not cut-and-
dried, and that the reader will discover some overlap.
The first subdivision of research I have labeled "International Events and
Foreign Policy Decision Making." The research I have called "International
Events" attempts to uncover patterns in the behavior of nations by utilizing
AI techniques developed to facilitate such a process through use of a
computer. These models are not designed to simulate how national authorities
came to produce such behavior; in that respect, they constitute a "minority
line" of research within the subfield. However, they can be viewed as efforts
to make the vast capabilities of the computer perform tasks similar to those
which international relations scholars perform when they endeavor to make
sense of sequences of international events. Schrodt's chapter uses pattern
6 Valerie M. Hudson
recognition techniques to tease out similarities and dissimilarities in sequences

of events composing international crises. Chapter 8 posits a performance-
oriented rule-based production system enabling one to postdict to an events
data set.
The "Foreign Policy Decision Making" research consists of work aimed
at modeling the information-processing and policy decision-making activities
of political actors. Brian Job and Doug Johnson explicate their UNCLESAM
program, a rule-based simulation, which captures the basic structures and
processes of U.S. decision making toward Latin America. Donald Sylvan
and his colleagues detail JESSE-a model of Japanese energy policy making
that incorporates the theory of generic tasks in its architecture. JESSE's
design permits it to function as a compiled reasoning system. Margee Ensign
and Warren Phillips discuss the application of AI/IR methods to understanding
the development policy choices of Third World leaders. In addition to assisting
them in this primary task, the authors note that the use of AI/IR methods
permits them to envision their model as a classroom teaching tool as well.
The second research subsection, "The Discourse of Foreign Policy,"
involves the search for techniques to interpret, process, and critically analyze
textual material of relevance to the study of international relations. G.R.
Boynton uses the notion of "interpretive triples" to discover how the Senate
Foreign Relations Committee attempted to make sense of the Reagan
administration's foreign policy during the summer of 1987. Sanjoy Banerjee
posits a "computational hermeneutic" model of Cold War superpower rhetoric
and shows how each side helped to "reproduce" the Cold War script of its
adversary. David Sylvan and his coauthors discuss how the notion of
"grounded theory" helps them to abduce categories in textual data consistent
with the self-understanding of those who generated the text. Finally, the
chapters by John Mallery and Gavan Duffy detail the components of the
RELATUS system introduced in the contribution by Hayward Alker et al.
in the conceptual section of the volume.
Grateful acknowledgments are due those who helped make this volume
a reality. Without the support of the David M. Kennedy Center for International
Studies at Brigham Young University and of Hayward R. Alker, Jr., this
volume probably would not exist.
Without the encouragement and advice of Donald Sylvan, I would never
have attempted this project. He and Philip Schrodt were crucial in supporting
me through the grimmer moments of editorship, and their advice proved
invariably sound. Brian Job was also very helpful. I would like to thank
John Mallery and Don Sorenson for helping to fill the gaps in my knowledge
on a variety of matters. Without the hard work of Louis Floyd, the manuscript
would never have assumed its final form. Several tough figures were rendered
camera-ready by Beccy Martin's selfless assistance. I am also appreciative
Introduction 7
of the financial support provided by the Department of Political Science at

Brigham Young University.
Finally, I would like to dedicate this volume to my grandmother, Roberta
Edstrom, whose memories made me whole and whose example lights my
way.
Notes
1. Computational modeling is the term to be preferred over AI modeling in this
regard, for the models put forth in this volume do not aspire to the ultimate goal
of AI, which is to produce a machine that can in some significant sense be said
to possess human intelligence. However, since most of the methodology used is
derived from AI, as long as the differentiation in objective is understood, I see no
reason not to make use of the more widely recognized acronym of "Al."
2. The term originates from the two edited volumes, both entitled Quantitative
International Politics, that came out during the heyday of this type of analysis in
international relations.
3. Whether the human in question be the researcher (see Schrodt, Mefford, Hudson
pieces this volume) or the research subject.
4. David Sylvan makes a strong case for a "possibilist" research agenda, and
this finds echoes in the works of others, as well, such as Hayward Alker, Stuart
Thorson, James Bennett, and Gavan Duffy.
Bibliography
Alker, Hayward, and C. Christensen, 1972. "From Causal Modelling to Artificial
Intelligence: The Evolution of a United Nations Peace-keeping Simulation," in J.A.
LaPonce and Paul Smoker (eels.), Experimentation and Simulation in Political
Science, Toronto: University of Toronto Press.
Harre, Rom, 1984. Personal Being, Cambridge, Massachusetts: Harvard University
Press.
Harre, Rom, and P.F. Secord, 1973. The Explanation of Social Behavior, Totowa,
New Jersey: Littlefield, Adams, & Co.
Hudson, Valerie M., 1987. "Using a Rule-Based Production System to Estimate
Foreign Policy Behavior: Conceptual Issues and Practical Concerns," in Stephen
Cimbala (ed.), Artificial Intelligence and National Security, Lexington, Massachu-
setts: Lexington Books, pp. 109-132.
Root-Bernstein, Robert, 1989a. "How Scientists Really Think," Perspectives in Biology
and Medicine, 32, 4, Summer, pp. 472-488.
___ , 1989b. Discovering, Cambridge, Massachusetts: Harvard University Press.
Schrodt, Philip, 1984. "Artificial Intelligence and the State of Mathematical Modeling
in International Relations." Paper presented at the U.S. and Swiss National Science
8 Valerie M. Hudson
Foundation Conference on Dynamic Models of International Conflict, Boulder,

Colorado, October 31-November 3.
Taylor, Charles, 1985. Philosophy and the Human Sciences, New York: Cambridge
University Press.
Weinberg, Gerald, 1975. An Introduction to General Systems Thinking, New York:
John Wiley and Sons.
1
Artificial Intelligence and
International Relations: An Overview
Philip A. Schrodt
Artificial intelligence techniques have been used to model international

behavior for close to twenty years. Alker and his students at MIT were
generating papers throughout the 1970s (Aiker and Christensen 1972; Alker
and Greenberg 1976; Alker, Bennett, and Mefford 1980), and by the the
early 1980s work at Ohio State was responsible for the first article to apply
AI in a mainstream IR journal (Thorson and Sylvan 1982) and the first
edited book containing a number of AI articles (Sylvan and Chan 1984).
This volume provides the first collection of essays focusing on AI as a
technique for modeling international behavior, 1 an approach commonly, if
controversially, labeled "AI/JR." The essays come from about twenty-five
contributors at fifteen institutions across the United States and Canada; the
substantial foci range from Japanese energy security policy to Vietnam
policy in the Eisenhower administration.
The purpose of this overview is twofold. First, it will provide a brief
background of relevant developments in AI in order to provide some
perspective on the concepts used in AI/IR. Second, it will identify some
common themes in the AI/IR literature that use artificial languages for
modeling. I will not deal with any of the issues in depth, nor provide
extensive bibliographical guidance, as these are ably presented in the chapters
themselves (e.g., those by Mefford and Purkitt in this volume; Mallery [1988)
also provides an excellent introduction). I will not discuss the natural language
processing (NLP) literature-which is covered in the chapters by Mallery,
Duffy, and Alker et al.-though I will discuss projects that use text as a
source of information for development of models rendered in artificial language
This research was supported in part by National Science Foundation Grant SES-8910738 and
by the University of Kansas General Research Allocation 3884-XQ-0038. My thanks to John
fllallery for helpful comments on an earlier draft.
9
10 Philip A. Schrodt
(e.g., Bennett and Thorson; Boynton; Sylvan, Milliken, and Majeski). This
chapter does not purport to provide a definitive description of the AI/IR
field; it is simply one person's view of the organization of the field at the
moment. In contrast to many other modeling approaches, the AI/IR com-
munity is characterized by a healthy level of internal debate. This chapter
is the overture, not the symphony. I intend only to draw your attention to
themes; the details, in both melody and counterpoint, are found in the
chapters that follow.
The label "artificial intelligence" is, ironically, rejected by a majority of
the authors in this volume as a description of their shared endeavor. The
preferred label is "computational modeling," which acknowledges the field's
intellectual roots in the formal modeling and computer simulation literature
within political science, rather than in the AI literature of computer science.
As will be noted below, the AI/IR efforts utilize only a tiny subset of AI
methods, and in many respects Al/IR overlaps at least as much with cognitive
psychology as with computer science.
The AI label poses two additional problems. The most severe is guilt by
association with "the AI hype": the inflated claims made for AI by the
popular media, science fiction, and consulting firms. The AI hype has been
followed by the backlash of the "AI winter," and so AI/IR risks being caught
in a counterrevolution just as it is beginning to produce results.
The second problem is the controversial word "intelligence." In the AI
hype, "intelligence" has usually been associated with superior intelligence
such as that exhibited by Star Wars robots (either the George Lucas or
Ronald Reagan variety). The most common retort I encounter when presenting
AI/IR overviews to unsympathetic audiences is: "You can't model politics
using artificial intelligence; you'd have to use artificial stupidity. " 2 As the
chapters that follow indicate, that is precisely our shared agenda! "Artificial
stupidity" involves limited information processing, heuristics, bounded ra-
tionality, group decision processes, the naive use of precedent and memory
over logical reasoning, and so forth. These features of human reasoning,
amply documented in the historical and psychological literature, are key to
AI/IR but largely absent from optimizing models of the dominant formal
paradigm in political science, rational choice (RC). Ironically, the true
"artificial" intelligence is utility maximization, not the processes invoked in
computational models.
All this being said, one must confront two social facts. First, the term
"computational modeling" has not caught on because it is not reinforced
by the popular media. Second, Al/IR has borrowed considerably from that
part of computer science and the cognitive sciences which calls itself
AI and International Relations: An Overview 11
"artificial intelligence," including the widespread use of LISP and Prolog as

formal languages, the formalization of rules, cases and learning, and a great
deal of vocabulary. In the spirit of mathematician David Hilbert's definition
of geometry as "that which is done by geometers," the AI label will probably
stick.
AI in the Early 1980s

The term "artificial intelligence" refers to a very large set of problems
and techniques ranging from formal linguistic analysis to robots. Researchers
in "AI" may be mathematicians or mechanics, linguists or librarians, psy-
chologists or programmers. Schank (1987:60) notes:
Most practitioners would agree on two main goals in Al. The primary goal is
to build an intelligent machine. The second goal is to find out about the nature
of intelligence. . . . [However,] when it comes down to it, there is very little
agreement about what exactly constitutes intelligence. lt follows that little
agreement exists in the AI community about exactly what AI is and what it
should be.
Research in AI has always proceeded in parallel, rather than serially, with

dozens of different approaches being tried on any given problem. As such,
AI tends to progress through the incremental accumulation of partial solutions
to existing problems, rather than through dramatic breakthroughs. None-
theless, from the standpoint of Al/IR, there were two important changes
in AI research in the late 1970s.
First, rule-based "expert systems" were shown to be able to solve messy
and nontrivial real-world problems such as medical diagnosis, credit approval,
and mechanical repair at the same level of competence as human experts
(see, for example, Klahr and Waterman 1986). Expert systems research
broke away from the classical emphasis in AI on generic problem solving
(e.g., as embodied in chess-playing and theorem-solving programs) toward
an emphasis on knowledge representation. Expert systems use simple logical
inference on complex sets of knowledge, rather than complex inference on
simple sets of knowledge. The commercial success of expert systems led
to an increase in new research in AI generally-the influx of funding helped-
and spun off a series of additional developments such as memory-based
reasoning, scripts, schemas, and other complex knowledge representation
structures.
Second, the personal computer, and exponential increase in the capabilities
of computers more generally, brought the capabilities of a 1960s mainframe
onto the researcher's desk. The small computers also freed AI researchers
from dependence on the slow and idiosyncratic software development designed
for centralized mainframes. The "AI style" of programming led to a generation
of programmers and programming environments able to construct compli-

cated programs that would have been virtually impossible using older
languages and techniques.
All of this activity lead to a substantial increase in the number of people
doi~g AI. The American Association for Artificial Intelligence (AAAI) was
founded in 1979, had 9,935 members by 1985 and 14,269 by 1986-a
growth of 43 percent in a single year. In short, AI in the 1980s was
accompanied by a great deal of concrete research activity in contrast to
faddish techniques such as catastrophe theory.
The AI Hype
Perhaps predictably, the increase in AI research was accompanied (and
partially fueled) by a great deal of hype in the popular and semiprofessional
media. An assortment of popular books on AI were produced by researchers
such as Feigenbaum (Feigenbaum and McCorduck, 1983; Feigenbaum and
McCorduck and Nii, 1988), Minsky (1986), and Schank (Schank and Riesback,
1981); at times these reached sufficient popularity to be featured by paperback
book clubs. Journalists such as McCorduck (1979), Sanger (1985), and
Leithauser (1987) provided glowing appraisals of AI; these are only three
of the hundreds of books and popular articles appearing in the early to
mid-1980s. Concern over the Japanese "Fifth Generation Project" (Feigen-
baum and McCorduck, 1983) provided impetus for the wildly unrealistic 3
"Strategic Computing Initiative" of the Defense Advanced Research Projects
Agency (DARPA, 1983).
These popular works provided a useful corrective to the outdated and
largely philosophical criticisms of Dreyfus (1979) and Weizenbaum (1976)
about the supposed limits of Al. By the early 1980s researchers had made
substantial progress on problems that by any reasonable definition required
"intelligence" and were exhibiting performance comparable to or exceeding
that of humans. However, the popularizations were understandably long on
concepts and short on code, and their explicit or implicit promises for
continued exponential expansion of the capabilities of various systems did
not take into account the tendency of technological innovation to follow a
logistic curve. 4 Because the promises made in these popularizations were
based largely on laboratory results that had not been scaled up nor widely
applied in real-world settings, such promises set up AI for a fall.
The AI Winter
The hype of the mid-1980s leveled off by the latter part of that decade
and some segments of the AI community-particularly companies producing
specialized hardware-experienced the "AI Winter." However, the decline of
AI was more apparent than real and reflected the short attention span of
the popular press as attention turned away from AI to global warming,

superconducting supercolliders, parallel processing, and cold fusion. Exper-
imental developments, most notably neural networks, continued to attract
periodic media attention, but the mainstream of AI assumed a level of
glamor somewhere between that of biotechnology and X-ray lasers: yesterday's
news, and somewhat suspect at that.
Yet ironically, the well-publicized bankruptcies of "AI firms" (see, for
example, Pollack 1988) were due to the success rather than the failure of
Al. As AI techniques moved out of the laboratories and into offices and
factories, commercial demand shifted from specialized "AI workstations"
and languages such as LISP to systems implemented on powerful general-
purpose microcomputers using standard procedural programming languages
such as C or off-the-shelf expert systems shells. AI research moved in-
house and was diffused into thousands of small applications rather than a
few large ones.
Overall, the AI field remained very healthy. Although membership in the
AAAI declined in 1988 and 1989, dropping to 12,500 members, the 1989
International Joint Conference on Artificial Intelligence, the AI equivalent of
the International Political Science Association, was large enough to require
the Detroit Convention Center, fill every convention hotel in downtown Detroit
and nearby Windsor, Ontario, and all this despite a $200 conference registration
fee.
Beyond the issues of popular perception, it is important to note that the
future of Al/IR is largely independent of the successes or failures of AI
generally. Whether a chess-playing program will be able to defeat the reigning
human grand master or whether simultaneous translation of spoken language
is possible will have no effect on most AI/IR research. Even if mainstream
AI has some implications for the development of computational models of
international behavior, the Al/IR literature is primarily shaped by literatures
in psychology, political science, and history rather than computer science.
The techniques borrowed from computer science are only tools for imple-
menting those theories.
AI/IR Research: A Framework

This section will attempt to structure the various sets of problems studied
in AI/IR. For example, there has been frequent confusion outside the field
as to why discourse analysis (represented in this volume by Boynton; Thorson
and Bennett; and Sylvan, Milliken, and Majeski) should have anything to
do with rule-based models (e.g., Job and Johnson) because the techniques
are entirely different. The simple answer is that the AI/IR literature is
primarily linked by underlying theories and questions rather than by meth-
odology. Although this is consistent with classical Kuhnian notions of science,
it is decidedly uncharacteristic of formal approaches to the study of

international behavior such as correlational analysis, arms-races models, and
game theoretic models of war initiation, which are largely linked by technique.
As noted earlier, any effort to find common themes in a field as conceptually
rich and disputatious as AljiR is fraught with the risk of oversimplification:
This chapter is simply a survey of the high points. AI/IR developed in an
evolutionary fashion; the organization I have presented below is a typology
imposed, ex post facto, on an existing literature, rather than an attempt to
present a consensus view of where the field is going.
The typology consists of three parts. The first category is the research
on patterns of political reasoning, which provides the empirical grounding
for models of organizational decision making. The second category involves
the development of static models of organizational decision making, which
aim to duplicate the behavior of an organization or system at a specific
point in time. This is the largest part of the literature in terms of models
that have actually been implemented, and it relies on the expert systems
literature in the AI mainstream. The final category contains dynamic models
that incorporate precedent, learning, and adaptation, which can show how
an organization acquired its behavior as well as what that behavior is. These
models are necessarily more elaborate and experimental, though some large-
scale implementations exist, notably the JESSE model of Sylvan, Goel, and
Chandrasekaran.
Patterns of Political Reasoning

The Psychological Basis. Virtually all work in AI/IR acknowledges an
extensive debt to experimental work in cognitive psychology. Although a
wide variety of approaches are cited, two literatures stand out.
The first influence is the work of Alien Newell and Herbert Simon on
human problem solving (Newell and Simon 1972; Simon 1979, 1982). Simon's
early work pointing to the preeminence of satisficing over maximizing
behavior is almost universally accepted in Al/IR, as is the Neweii-Simon
observation that human cognition involves an effectively unlimited (albeit
highly fallible) memory but a fairly limited capacity for logical reasoning.
These assumptions about human problem solving are exactly opposite those
of the rational choice approach, where cognition involves very little memory
but optimization is possible. In addition to these general principles, other
work by Newell and Simon on specific characteristics of human problem
solving is frequently invoked, for example the re-use of partial solutions
and the distinction between expert and novice decision making.
The second very large experimental literature is the work of Daniel
Kahneman, Paul Slovic, Amos Tversky (KST), and their associates in exploring
the numerous departures of actual human decision making from the char-
acteristics predicted by utility maximization and statistical decision theories

(Kahneman, Slovic, and Tversky 1982). This work has emphasized, for
example, the importance of problem framing, the use of heuristics, the
effects of familiarity and representativeness, and so forth.
Even though the experimental results, general principles, and concepts
of these two literatures are used extensively in Al/IR, their theoretical
frameworks are not. Newell and his students have developed a general
computational paradigm for AI, SOAR (Laird, Rosenbloom, and Newell 1986;
also see Waldrop 1988), but to my knowledge it has not been applied in
the IR context. "Prospect theory," the term usually applied to the KST
work, is also rarely used. These research results are used instead to explicate
some of the characteristics of IR decision making, which, because it is
organizational and frequently involves unusual circumstances such as the
decision to engage in lethal violence, is far removed from the individualistic
studies of much of the psychological literature. Some work on group decision-
making dynamics has also been used-for example, Pennington and Hastie
on decisions by juries (cited in Boynton)-but in general this literature is
smaller and less well known in cognitive psychology than the theories and
experiments on individuals.
Knowledge Representation. Consistent with the Neweii-Simon approach,
AI/IR models are heavily information-intensive. However, the theory of data
employed in AI/IR is generally closer to that of history or traditional political
science than it is to statistical political science. Most of the chapters in
this volume use archival text as a point of departure; the remainder use
secondary data such as events that were originally derived from text using
procedures similar to content analysis. The ordinal and interval-level measures
common to correlational studies, numerical simulations, and expected utility
models are almost entirely absent.
This, in turn, leads to the issue of knowledge representation, which is
a theme permeating almost all of the papers and which accounts for much
of their arcane vocabulary and seeming inconsistency. Whereas behavioral
political analysis essentially has only three forms of knowledge represen-
tation-nominal, ordinal, and interval variables-At presents a huge variety,
ranging from simple if . . . then statements and decision trees to scripts
and frames to self-modifying programs and neural networks. This surfeit
of data structures is both a blessing and a curse. It provides a much broader
range of alternatives than are available in classical statistical or mathematical
modeling, and certainly provides a number of formal structures for repre-
senting the large amounts of information involved in political decision making;
this comes at the expense of a lack of closure and an unfamiliar vocabulary.
Part of this problem stems from the fact that knowledge representation
concepts have yet to totally jell within their parent discipline of computer
science. In this regard AI/IR is quite different than the behaviorialist adoption
of statistical techniques and the RC adoption of economic techniques: In

both cases stable concepts and vocabulary were borrowed from more mature
fields.
Structure of Discourse and Argument. For the outsider, perhaps the most
confusing aspect of the AI/IR literature is the emphasis on the analysis of
political argument and discourse. This type of analysis is found in the articles
by Boynton; Sylvan, Milliken, and Majeski; and Bennett and Thorson; more
sophisticated tools for dealing with discourse are found in the natural language
processing (NLP) articles. At first glance, rummaging through the Congres-
sional Record or using Freedom of Information Act requests to uncover
Vietnam-era documents is the antithesis of the formal modeling approach:
Archival sources are the stuff of history, not models.
In fact, these analyses are at the core of modeling organizational cognition.
As such the archival work is simply specialized research along the lines of
the general psychological studies. The political activities modeled in AI/IR
are, without exception, the output of organizations. Because the AI/IR
approach assumes that organizations are intentional and knowledge-seeking,
it is important to know how they reason. Conveniently, organizations leave
a very extensive paper trail of their deliberations. Although archival sources
do not contain all of the relevant information required to reconstruct an
organization's behavior-organizations engage in deliberations that are not
recorded and occasionally purposely conceal or distort the records of their
deliberations-it is certainly worthy of serious consideration. 5
Static Modeling: Rule-Based Systems

Rule-based systems (RBS) are currently the most common form of AI/
IR model, and even systems that go well beyond rules, such as the JESSE
simulation, contain substantial amounts of information in the form of rules.
Contemporary RBS are largely based on an expert systems framework, but
the "production systems" that dominated much of AI modeling from the
late 1950s to the early 1970s are also largely based on rules; early production
system models of political behavior include Carbonell (1978) in computer
science and Sylvan and Thorson (1982) in JR. In addition to chapters by
the authors in this volume, other models of international behavior using the
rule-based approach have included Soviet crisis response (Kaw 1989), Chinese
foreign policy (Tanaka 1986), the political worldview of Jimmy Carter (Lane
1986) and Chinese policy toward Hong Kong (Katzenstein 1989).
In its simplest form an RBS is just a large set of if . . . then statements.
For example, a typical rule from Job and Johnson's UNCLESAM program-
a simulation of U.S. policy toward the Dominican Republic-has the form
IF U.S. Posture to the Dominican Republic government > 4

and
Stability Level > = 5

and
Stability Level Change > 0
THEN
Increment U.S. Use of Force Level by 1
An RBS may have hundreds or thousands of such rules; they may exist
independently, as in Job and Johnson or production system models, but
more typically are organized into hierarchical trees (for example, Hudson in
this volume or Kaw 1989). Typical commercial expert systems used for
diagnosis or repair have about 5,000 rules; most AI/IR systems are far
simpler. Mefford's chapter in this volume describes in considerable detail
RBS developments beyond basic if . . . then formulations; one should also
note that the boundaries between the more complicated RBS and other
types of models (for example, case-based reasoning and machine learning
systems) are extremely fuzzy. Nonetheless, virtually all AI models encode
some of their knowledge in the form of rules. 6
Despite the near ubiquity of rules in computational models, this approach
stands in clear contrast to all existing formal modeling traditions in political
science, which, without exception, use algebraic formulations to capture
information. These methods encode knowledge by setting up a mathematical
statement of a problem and then doing some operations on it (in RC models,
optimization; in statistics, estimation; in dynamic models, algebraic solution
or numerical simulation). The cascading branching of multiple rules found
in RBS is seldom if ever invoked; when branches are present they usually
only deal with boundary conditions or bifurcations7 and are simple in structure.
Although much of the impetus for the development of RBS in political
science came from their success in the expert systems literature, rules are
unusually well suited to the study of politics, as much of political behavior
is explicitly rule-based through legal and bureaucratic constraints. Laws and
regulations are nothing more than rules: These may be vague, and they
certainly do not entirely determine behavior, but they constrain behavior
considerably. Analyses of the Cuban Missile Crisis, for example, repeatedly
observe that the military options were constrained by the standard operating
procedures of the forces involved.
Informal rules-"regimes" or "operational codes" in theIR literature (e.g.,
Krasner 1983; George 1969)-impose additional constraints. For example,
in the Cuban Missile Crisis, John F. Kennedy did not consider kidnapping
the family of the Soviet ambassador and holding them hostage until the
missiles were removed, though in some earlier periods of international history
(e.g., relations between the Roman and Persian empires, circa 200 c.E.) this
would have been considered acceptable behavior.
In short, rule-based political behavior is not an "as if" proposition: It

can be empirically confirmed. However, because bureaucracies do not solely
follow their rules-in fact most bureaucracies would be paralyzed if they
attempted to do so-the extent to which rules can capture actual behavior
and the complexity required to do so is an open question. As the chapters
in this volume and other RBS research indicate, it is clearly possible to
simulate political behavior using rules. The complexity of these systems,
though substantially greater than the complexity of most existing formal
models (other than simulations) is also well within the limits of existing
RBS developed in other fields.
Despite the widespread use of rules in AI/IR models, many of the chapters
in this volume argue against rule-based formulations, or at least indicate
problems with the use of rules. This should not be interpreted as a rejection
of any use of rules but only as a rejection of depending solely on rules.
This, in turn, is a reaction to theoretical issues in AI rather than political
science: A variety of approaches advocate rules in some form as a universal
standard of knowledge representation and argue that all work in AI should
use a single, unified concept of knowledge representation and manipulation.
This comprehensive approach is rejected by almost all of AI/IR work as
premature at best, given the evidence from the cognitive and organizational
decision-making literature.
Dynamic Modeling: Adaptation and Learning

Learning is one of the most basic characteristics of human cognitive
behavior but has been largely absent from existing formal models in political
science.8 One of the most distinctive-and potentially revolutionary-char-
acteristics of the AI/IR models is the use of learning as a dynamic element.
By attempting to model not only what organizations do but why they do
it-in the sense of providing an explanation based on the prior experience
of the organization-these models can potentially provide greater detail and
process validity than those currently available.
An assortment of learning schemes are currently under development, but
most involve at least two elements. First, the basic rule of learning is
"Bureaucracies do not make the same big mistake twice." Reactions to a
situation are based in part on precedents, and the success or failure of a
particular plan will affect whether it is used in the future. Knowledge is
modified, not merely acquired.
The second general characteristic is the reapplication of solutions that
worked in the past to problems encountered in the present: In the JESSE
simulation this is called "compiled reasoning." As Mefford notes, this same
concept is central to Polya's scheme of human problem solving; more
recently it has been the central focus of the Laird, Rosenbloom, and Newell
SOAR project. The content of an organization's compiled reasoning in turn

depends on the problems it has previously encountered, and so the history
of the organization is important in determining what it will do.
This emphasis on learning and adaptation provides the "path dependence"
mentioned in a number of the chapters. International politics is not like a
game of chess where future plays can be evaluated solely on the basis of
the current board position. Instead, how a situation came into being may
be as important as its static characteristics. Deductive reasoning from first
principles alone is not sufficient. Learning also partially resolves the problem
of underdetermination in satisficing-it indicates not just the suboptimal
character of decisions but also substantially narrows which suboptimal
decision will be made.
Learning is such a basic human activity that it is taken for granted in
the informal bureaucratic politics literature, though in recent years a number
of studies reemphasizing the importance and problems of learning as a
separate activity have emerged (see, for example, Neustadt and May 1986
on precedent and analogy; Etheredge 1985 on learning in foreign policy;
Margolis 1987 on patterns). The difference is that the AI/IR models have
succeeded in formally modeling this learning process in a complex envi-
ronment, which no other models have. 9 To the extent that learning and
adaptation are key human behaviors, this is a major step forward.
Perhaps because learning is so inherently human, the extent to which it
can be simulated is frequently underestimated, if not rejected outright. One
of the most popular-and most inaccurate-characterizations of computer
models is "A computer can't do anything it wasn't programmed to do."
Strictly speaking, this may be true: A computer can't do in five minutes
anything a human with a pencil and paper couldn't do in eleven centuries
of sustained effort, 10 but for practical purposes there is a qualitative difference.
As a complex system, a computer can very easily be programmed to
perform in ways unanticipated by its programmers. By that same token
one can reject another hoary characterization, "Computers can't be creative."
In fact, there are a variety of rather straightforward techniques by which
programs can "create" solutions to problems that were unanticipated by
their programmers, and the creation of such programs has been a long-
standing research tradition in mainstream Al.
Modeling organizational learning involves modeling at least two different
types of learning. The first, and simpler, learning is knowledge acquisition:
the process by which precedents, analogies, cases, plans, or whatever are
acquired and retained. The second and more difficult phase is modeling the
adaptation of the means by which those are invoked. In other words, the
process of "reasoning by analogy" involves both the availability of analogous
situations and also the means by which the analogy is made: Neither issue
is self-evident.
Consider, for example, the situation of a failure of reasoning by analogy

in the Bay of Pigs invasion. This exercise was modeled on the earlier
successful CIA overthrows of Mossadegh in Iran and Arbenz in Guatemala.
Had the Marines been involved in the planning, analogies might have been
made to the problems of invading islands encountered in World War 11,
leading, presumably, to greater skepticism or at least better preparation.
Hence one could say that the Bay of Pigs failed because only "overthrow"
precedents were used to the exclusion of "invasion" precedents.
Alternatively, one could argue that the situation itself had been incorrectly
understood in the sense that important variables were not considered: The
theory that identified the precedents was in error. If the prevailing theory
indicates that two things should match and they did not, the organization
will recognize that something is wrong with the theory and change it.
Arguably, this occurred historically: The failure at the Bay of Pigs was
attributed to the lack of sufficient force and a clear U.S. commitment, and
therefore the next two U.S. interventions in the Third World-in the Dominican
Republic and Vietnam-involved overt use of large numbers of troops.
These two types of learning have very different implications for the
structure of the model, and the question as to which dominates is an
empirical one. Modeling the use of precedent is relatively straightforward,
but organizational modifications to the interpretation of precedent may be
more important, particularly when dealing with unexpected behaviors.
Precedent. The concept of precedent is probably second only to that of
"rules" in the discussions in this volume, and precedent is closely linked
to the issue of memory. In the simplest form, actors will simply do what
they've done before, and the best predictor for behavior in a complex system
is the patterns of history. Precedent, analogy, and case studies have strong
antecedents in the traditional political literature (e.g., Neustadt and May
1986), as well as being formally invoked in Anglo-American legal reasoning.
As with the use of rules, precedent does not require "as if" apologies: lt
is used openly and explicitly.
The use of precedent is, however, somewhat ambiguous. In foreign policy
discourse, precedent is most likely encountered as a justification-in other
words, it is invoked as an empirical regularity. In legal reasoning and in
much of the case-based reasoning literature, in contrast, it is a plan-a
series of actions which should be taken. In the latter sense a precedent is
merely a complex antecedent clause of an if ... then clause with a complex
consequent. These two uses of precedent are not mutually incompatible-
in a sufficiently regular and well-defined system, the use of precedents as
plans would cause them to become empirical regularities-but they are
different.
As noted in several of the chapters, empirical studies of actual policy
deliberations find little evidence of a dominant role for precedent. Several
factors may account for this. First, in contrast to the legal arena, the
international system is neither well defined nor particularly regular, and
consequently the precedents are not necessarily clear. Furthermore, in IR
discourse, a precedent such as "Munich," "Pearl Harbor," or "Vietnam" is
most likely to be invoked as something to be avoided, not to be implemented. 11
Precedents used repeatedly are incorporated into the standard operating
procedures of the organization-they become rules. Consequently, precedent
may be a powerful tool for decision making even if it isn't actively invoked
in debate. The second possible problem is that precedents are probably
generalized into "ideal cases." If a decision-maker refers to the danger of
a coup in El Salvador, what is usually invoked is not a specific coup 12 but
rather coups in general.
Compiled Reasoning and Structure. On the surface, compiled reasoning
is only a modest enhancement of existing psychological models: The notion
that individuals and organizations reuse prior successful behavior is not
particularly controversial and is certainly strongly supported by experimental
evidence with individuals going back several decades. However, the critical
contribution of such models may be in the specification of "learning-driven
models" (LDMs)-models whose key dynamics are determined by learning.
In particular, this might begin to concretize the heretofore exceedingly mushy
concept of "structure" in political behavior.
For example, suppose that at any given time, the actors in a system
can be viewed (or modeled) as simple stimulus-response actors-ceteris
paribus, given an input, one can predict the resultant behavior. This, in
turn, provides a set of mutual constraints on those activities, which we
refer to as "structure." For example in the 1970s, any Eastern European
state, on the basis of Soviet activities in Berlin, Hungary, and Czechoslovakia
during the 1950s and 1960s could reasonably assume that excessive economic
and political liberalization would lead to Soviet military intervention. From
the standpoint of the actor, this is simply a rule of the system.
However, in an LDM, every event has the potential of changing those
reaction functions-in other words, the system is self-modifying (either
through accumulation of cases or modification of rules). Although most
events do not change the functions, when learning occurs, the reaction
functions of the system may change dramatically. For example, by 1989,
Eastern European states reacted as if economic and political liberalization
would not cause Soviet intervention after some initial experimentation by
Poland along these lines was reinforced. Because some-if not most-of
the knowledge of structure is embodied in complex qualitative structures,
the change is not necessarily incremental, and it may be mutually reinforcing
(as happened, for example, in Hungary, Poland, and the GDR in the autumn
of 1989, followed later by Czechoslovakia, Bulgaria, and Romania) so as to
cause a series of fairly dramatic changes.
This does not, however, mean that the situation is chaotic, a key advantage
of LDMs over the other formulations. Most of the knowledge base and the
learning mechanism has not changed;13 only the output has changed. The
environment, the actors, and the parameters of their decision making are
almost unmodified; the change is embodied in relatively simple and predictable
modifications of knowledge and rules for interpreting that knowledge, often
empirically available in the archival record. Realistically modeling this type
of behavior is a difficult task, and the AlfiR system that comes closest to
embodying it, unsurprisingly, is the complex JESSE simulations. But this
objective underlies many of the chapters.
The Upshot: The Ideal Model

To summarize these points, the following is a list of the characteristics
I believe most researchers would include in an ideal AlfiR model.
• The model would be consistent with the psychological literature on

individual and group decision making, in particular it would involve sub-
optimal reasoning, heuristics, anchoring, and other aspects of "artificial
stupidity."
• Both the actions predicted by the model and the reasoning reflected
in the determination of those actions would not be inconsistent with
that found in actual political debate.
• The model would store knowledge in complex, qualitative data structures
such as rules, scripts, frames, schemas, and sequences. Among that
knowledge, though not necessarily dominant, would be precedents and
cases.
• The model would learn from both successes and failures: Successful
solutions would be likely to be reused; unsuccessful solutions or ob-
servations about changes in the environment would cause changes in
the rules that are used to solve problems. This learning would be similar
to that seen in actual political behavior.
• The overall model would provide a general engine for the study of
international politics-in other words, it would work on a variety of
problems. The specific knowledge required to implement the model, of
course, would differ significantly with the problem.
None of the existing models incorporates all of these factors, and none
is considered even close to a final solution to them. However, almost all of
the chapters in this volume contribute to this agenda, and quite a number
of the components have already been demonstrated.
AI and Contemporary Political Science

One of the most intriguing aspects of the Al/IR research is the breadth
of the substantive foci of these studies. The topics dealt with in this volume
include the Johnson administration's Vietnam policy, Dwight Eisenhower's
Vietnam policy, Japanese energy security policy, development, U.S. policy
in the Caribbean, international relations during the early Cold War period,
the international behavior recorded in the BCOW and CREON data sets,
the Senate Foreign Relations Committee on the Persian Gulf in the 1980s,
the Soviet intervention in Hungary, and Luttwak's theory of coups d'etat.
This list reads like the topics of any randomly chosen list of articles on
international relations, in distinct contrast to the arms race and world
modeling literatures, which have focused on fairly narrow types of behavior.
By that same token, there is a strong empirical focus in virtually all of the
studies. The research is grounded in the detailed study of actual political
behavior, not in techniques borrowed from economics, computer science,
or mathematics.
AI and Rational Choice

The most important alternative model to that proposed in the Al/IR
approach is the rational choice approach, which presumes that political
behavior, like economic behavior, can best be modeled by assuming individuals
optimize their choices within a system of preferences and constraints without
any higher cognitive processes. RC uses, for the most part, expected utility
decision making and game theory, and is important both because of its
dominant role in explaining domestic behavior and increasing application as
a theory of international behavior. 14 In addition to its role as "straw man
of choice" in contemporary discourse in political science, RC provides an
alternative to Al/IR in positing an explicit cognitive mechanism, in contrast
to statistical studies and most dynamic models, which merely posit regularities.
This debate between RC and cognitive psychology is not confined to IR but
is found generally in the social sciences: Hogarth and Reder (1987) provide
an excellent survey of the arguments; much of the KST agenda is directed
toward falsifying RC assumptions; Simon (1985) deals explicitly with these
issues in terms of political science.
Although in general AI/IR is viewed as competition to RC-and this is
the position taken by most authors who address themselves to the issue-
the possibility remains that the two approaches are complementary: Problems
appropriate for an RC framework might be inappropriate for AI and vice
versa. Most of the discourse analysis and NLP articles would probably concur
with this characterization, and NLP tends not to address the RC issue at
all. Using AI in a situation that can be accurately described by optimization
using preferences and constraints would be using a sledgehammer to kill

a fly: The problem of nuclear deterrence comes to mind. In relatively static
situations of reasonably complete information, limited bureaucratic infighting,
quantitative or strictly ordered payoff's, and repeated plays, the expected
value framework may be completely appropriate.
This assumption of complementarity is the position of many economists
who are now using computational modeling: They maintain the basic RC
concepts and vocabulary but use computational models to deal with phe-
nomena such as memory and learning that confound existing mathematical
techniques. If one can take the structuring of a situation as known and
static-for example, the voting decision in a stable liberal democracy doesn't
include the option of changing the rules of voting-then RC will predict
behavior quite efficiently. However, when self-modification is an option, simple
RC models aren't very useful.
Alternatively, the two approaches may be simply competitive; this view
is implicit in most of the chapters in this volume that discuss RC. Fundamental
to virtually all critiques of RC is criticism on empirical grounds: The "as
if" approach, though possibly acceptable when formulated by Milton Friedman
in 1953, is unacceptable when viewed in the light of thirty years of
experimentation that has failed to validate those assumptions except under
highly artificial settings. The AljiR literature, with its base in experimental
psychology and its heavy use of primary data, provides an alternative
formulation of political rationality with a superior empirical grounding.
In evaluating this debate, note that the difference between AI/IR and RC
does not lie in the assumption of preferences and goal-seeking behavior.
The crucial difference is in the issue of optimization: RC assumes utility
maximizers; AI/IR assumes individuals and organization have inadequate
information-processing abilities to optimize. In addition, AI/IR generally
assumes a much more complicated and dynamic world than RC theories,
which it can do because the models are computational rather than algebraic.
For example, RC models usually assume that the consequences of actions
are known (with a known degree of uncertainty), whereas AI/IR approaches
may try to model the acquisition of that knowledge. AI/IR, following KST,
postulates that the framing of a problem is a very important phase of
finding the answer and, in fact, may well determine the answer. Following
the Simon-Newell tradition, memory is not considered a constraint either
in the cognitive theory or the computational model, so huge amounts of
information in complex structures can be used. In this sense, AI/IR and
RC are mirror images: AI assumes sophisticated memory but limited pro-
cessing; RC assumes sophisticated processing (optimization) but limited
memory.
On the down side, AI has far less mathematical closure than RC. RC's
intellectual coherence stems in no small part from a series of mathematical
bottlenecks imposed by the available techniques. Common mathematical

tools such as fixed-point theorems and results from game theory severely
restrict the assumptions that can be made if a model is to be mathematically
tractable. On the one hand, this gives RC an "assume the can opener"
air; 15 on the other hand, it provides a large body of shared assumptions so
that many new results are widely applicable. RC has a cumulativeness that
is less evident in AI/IR. Because AI is algorithmic rather than algebraic,
there are virtually no constraints on the complexity of the underlying
argument; the techniques, in fact, tend to be data-limited rather than
technique-limited, a problem shared with numerical simulations, as Mefford
notes.
AI/IR and Data

The AI/IR approach is strongly empirical, another aspect that differentiates
it from the RC tendency to refine theoretical concepts with little concern
for empirical testing. From a sociology of science standpoint, the empirical
focus of AI/IR is unsurprising given its dual roots in experimental psychology
and the fact that most of the early researchers, from Alker onward, initially
studied politics using behavioralist statistical methodologies. The AI/IR
approach is developing, from the beginning, testable theories with the
corrective feedback empirical studies provide.
The data used in AljiR resembles that of historical/archival research
more than that of the quantitative statistical traditions of the 1960s and
1970s. The majority of the chapters in this volume use primary text or
interviews as their data source, and they also emphasize issues such as
context, structure, and discourse. Even those studies not using text (for
example, Hudson or Schrodt) utilize very detailed descriptive sequences with
thousands of points, a distinct contrast to the annualized forty-year time
series typical of contemporary statistical studies. Almost all existing political
research has concentrated on simple data structures, usually the rectangular
case/variable data array. Even when more complicated structures were
built-as, for example, in factor analysis or Guttman scaling-these were
done within the constraints of linear models. However, it seems obvious
that political behavior involves complex data structures such as rules and
hierarchies, scripts and sequences, plans and agendas. The rational choice
tradition has begun to work with some complex structures-for example,
with agenda setting-but has a very limited empirical tradition. In contrast,
most of the methods being developed in AI/IR are general: For example,
the methods of analyzing informal rules in the context of bureaucracies or
international regimes should be of substantial use to those studying formal
institutions such as Congress or the evolution of international regimes.
The question of what will be tested remains open. In general there are
two possible criteria: outcome validity and process validity. A model with
outcome validity would reproduce or predict behavior but make no pre-

sumptions that the mechanisms through which that behavior was generated
corresponded to those in the international system. The chapters using
standard IR data sets (e.g., Hudson and Schrodt) are most clearly in this
camp. A stricter, and more useful, criterion is process validity: The mech-
anisms by which an outcome is reached should also correspond to those
observed in the actual system. The models strongly based in archival and
primary source material (e.g., JESSE; Sylvan, Milliken, and Majeski) are
coming closer to achieving this. A problem in testing any model of process
validity, however, is the arbitrariness and incompleteness of the empirical
record: A model may produce behavior and rationales that are entirely
plausible (as evaluated by experts) but that are not found in the actual
historical record.
Conclusion
Despite the tendency at paradigm proliferation in the IR literature-two
articles (at most) seem to establish a new IR paradigm-the AI/IR literature
has most of the characteristics of a classical Kuhnian paradigm, including
a new set of questions, theories of data, and techniques. As a paradigm,
AI/IR allows one to look at old data in new ways; one could argue that it
also arose out of failures of the behaviorialist modeling techniques to deal
with the contextual complexity found in primary source material.
The projects reported in this volume are in various stages of completion.
Although the "trust me" assurances in research are quite rightly viewed
with skepticism in a new and intensely hyped technique such as AI, most
of this research has a large inductive component based in extensive primary
source material rather than data sets from the ICPSR and justifiably proceeds
slowly. The completed research, frequently of considerable complexity (e.g.,
Hudson, the JESSE simulation, Job and Johnson), is, I hope, a harbinger
for the results of the projects still in progress (e.g., Ensign and Phillips).
Its vocabulary is evolving and its critical concepts emergent, yet the quantity
and diversity of the AI/IR research projects have already transcended those
of most faddish modeling techniques. The research reported in this volume
opens a variety of doors to further research, and the list of potentially
interesting projects is far from exhausted.
Notes
1. Cimbala (1987) and Andriole and Hopple (1988) provide introductions oriented
toward U.S. defense concerns, but these are only a narrow part of the field of IR
generally.
2. This seemingly original joke seems to occur to almost everyone. . . .
3. For example, DARPA's 1983 timetable calls for the following developments by
1990: vision subsystems with "1 trillion Von-Neumann equivalent instructions per
second"; speech subsystems operating at a speed of 500 MIPS that "can carry on
conversation and actively help user form a plan [sic)," and "1,000 word continuous
speech recognition." Each of these projected capabilities is 10 to 100 times greater
than the actual capabilities available in 1990, despite massive investments by DARPA.
For additional discussion, see Waldrop (1984); Bonasso (1988) provides a more
sympathetic assessment.
4. A 10,000-rule expert system is unlikely to achieve ten times the performance
of a 1,000-rule system. To the contrary the 1,000-rule system will probably have
80-90 percent of the functionality of the large system; the additional 9,000 rules
are devoted almost exclusively to the residual 10-20 percent of the cases.
5. Conceptually, this effort is similar to the "cognitive mapping" methodology
pursued in Axelrod (1976); the "operational code" studies (e.g., George 1969; George
and McKeown 1985) are other antecedents.
6. Neural networks are the primary exception to this characteristic and are a
current research focus precisely because they offer an alternative to rule-based
formulations.
7. For example an action A would be taken in the expected utility formation
E(A) = p(100) + (1-p)(-50)
if and only if p > 1/3; the dynamic model
is stable if and only if la I< 1 and so forth.

B. Formal models of simple learning are common in the psychological literature,
but these are largely inappropriate to the organizational learning of complex, ill-
defined tasks common to political settings.
9. Learning is another aspect acquired from the AI tradition. One of the key
elements of LISP, the dominant AI programming language, is the absence of a
distinction between program and data. As a consequence programs could be "self-
modifying," changing to adapt to a problem at hand. This is the essence of learning,
and so models drawing on the AI tradition had a rich set of learning concepts at
its core, whereas learning was difficult to add to RC or dynamic models.
10. Figuring eight-hour days and a sustained rate of 100 calculations per hour
for the human; a very modest 1 MIPS for the computer.
11. This can occur for normative as well as pragmatic reasons: For example,
Allison (1971:197) emphasizes the strong negative impact the Pearl Harbor analogy
had on Robert Kennedy's assessment of the option of bombing Cuba during the
Cuban Missile Crisis. "I now know how Tojo felt when he was planning Pearl Harbor,"
Kennedy wrote during an ExCom meeting, and he would later write "America's
traditions and history would not permit . . . advocating a surprise attack by a very
large nation against a very small one."
12. Unless there is a clear and obvious precedent: The future of Ferdinand Marcos
in the Philippines was discussed in terms of Marcos as "another Somoza," referring
to the Nicaraguan dictator whose fall led to the establishment of the Sandinista
regime. Usually, however, the search for precedent does not go very deep.
13. In contrast, consider the treatment of international crises found in Snyder
and Diesing (1977), which uses an RC approach. For a given configuration of payoffs
at any stage in a crisis, Snyder and Diesing can analyze, using game theoretic
concepts, the likely behavior of the actors in the crisis (or, more likely, induce the
payoffs from the behavior). But this approach does not provide a means of moving
from one game matrix to the next: It predicts only a single decision, not a succession
of decisions. An LDM approach, in contrast, would seek to model the changes in
the payoffs as well as the decisions themselves. In a much simpler framework, the
"sequential gaming" models of the RC tradition are also trying to attack this problem
using algebraic methods.
14. Supporters of the rational choice approach frequently consider it to be as
pervasive in IR as in domestic politics. Although it has dominated one problem-
the counterfactual analysis of nuclear war and nuclear deterrence (see Brams 1985)-
it is a relatively recent newcomer in the remainder of IR, dating mostly from the
work of Bueno de Mesquita and his students, and more recently "crossovers" such
as Ordeshook and Niou (see Ordeshook 1989). Even this work deals primarily with
a single issue, war initiation. One need only compare the scope of the arms race
bibliography of Anderton (1985), with 200+ entries, or the dynamic simulation
literature (e.g., Guetzkow and Valdez, 1981) to see how relatively small the RC
literature is in JR. Outside of IR, of course, RC is clearly the dominant formal
paradigm in political science.
15. For the benefit of those few who do not understand this allusion, the underlying
joke goes as follows: A physicist, a chemist, and an economist were stranded on
a desert island. Exploring the island, they found a case of canned food but had
nothing to open it with. They decided to share their expertise and, being academics,
each gave a short lecture on how their discipline would approach the problem. The
physicist began, "We should concentrate sufficient mechanical force to separate the
lid from the can. . . . " The chemist beg~>n, "We should create a corrosive agent
to dissolve the lid from the can. . . . " The economist began, "First, assume we
possess a can opener. . . . "
Bibliography
Abelson, Robert P. 1973. "The Structure of Belief Systems," pp. 287-339. In R.C.
Schank and K.M. Colby (eds.), Computer Models of Thought and Language. San
Francisco: Freeman.
Alker, Hayward J., James Bennett, and Dwain Mefford. 1980. "Generalized Precedent
Logics for Resolving Security Dilemmas." International Interactions 7: 165-200.
Alker, Hayward J., and C. Christensen. 1972. "From Causal Modeling to Artificial
Intelligence: The Evolving of a UN Peace-Making Simulation," pp. 177-224. In
J.A. LaPonce and P. Smoker (eds.), Experimentation and Simulation in Political
Science. Toronto: University of Toronto Press.
Alker, Hayward J., and W. Greenberg. 1976. "On Simulating Collective Security
Regime Alternatives," pp. 263-306. In M. Bonham and M. Shapiro (eds.), Thought
and Action in Foreign Policy. Basel: Birkhauser Verlag.
Alker, Hayward J., and F. Sherman. 1982. "Collective Security-Seeking Practice

Since 1945," pp. 113-45. In D. Frei (ed.), Managing International Crises. Beverly
Hills: Sage.
Allison, Graham T. 1971. The Essence of Decision. Boston: Little, Brown.
Anderton, Charles H. 1985. "Arms Race Modelling: Categorization and Systematic
Analysis." International Studies Association, Washington.
Andriole, Stephen J., and Gerald W. Hopple. 1988. Defense Applications of Artificial
Intelligence. Lexington, MA: Lexington Books.
Axelrod, Robert, ed. 1976. Structure of Decision. Princeton: Princeton University
Press.
Bonasso, R. Peter. 1988. "What AI Can Do for Battle Management: A Report of the
First AAAI Workshop on AI Applications to Battle Management." AI Magazine
9,3: 77-83.
Bonham, G. Matthew, and Michael J. Shapiro. 1976. "Explanation of the Unexpected:
The Syrian Intervention in Jordan in 1970," pp. 113-141. In Robert Axelrod (ed.),
The Structure of Decision. Princeton: Princeton University Press.
Brams, Steven J. 1985. Superpower Games. New Haven: Yale University Press.
Carbonell, Jaime G. 1978. "POLITICS: Automated Ideological Reasoning." Cognitive
Science 2: 27-51.
Cimbala, Stephen. 1987. Artificial Intelligence and National Security. Lexington, MA:
Lexington Books.
Defense Advanced Research Projects Agency. 1983. "Strategic Computing: New
Generation Computing Technology: A Strategic Plan for its Development and
Application to Critical Problems in Defense." Memo, 28 October 1983.
Doyle, Jon. 1988. "Big Problems for Artificial Intelligence." A/ Magazine 9,2: 19-
22.
Dreyfus, Herbert L. 1979. What Computers Can't Do. New York: Harper and Row.
Etheredge, Lloyd S. 1985. Can Governments Learn? New York: Pergamon.
Feigenbaum, Edward A., and Pamela McCorduck. 1983. The Fifth Generation. Reading,
Mass: Addison-Wesley.
Feigenbaum, Edward A., Pamela McCorduck, and H. Penny Nii. 1988. The Rise of
the Expert Company. New York: Times Books.
George, Alexander L. 1969. "The 'Operational Code': A Neglected Approach to the
Study of Political Leaders and Decision-making." International Studies Quarterly
13: 190-222.
George, Alexander L., and Timothy J. McKeown. 1985. "Case Studies and Theories
of Organizational Decision Making." Advances in Informational Processing in
Organizations 2: 21-58.
Guetzkow, Harold, and Joseph J. Valdez, eds. 1981. Simulated International Processes:
Theories and Research in Global Modeling. Beverly Hills: Sage.
Hogarth, Robin M., and Melvin W. Reder, eds. 1987. Rational Choice: The Contrast
Between Economics and Psychology. Chicago: University of Chicago Press.
Kahneman, Daniel, Paul Slovic, and Amos Tversky, eds. 1982. Judgement Under
Uncertainty: Heuristics and Biases. Cambridge: Cambridge University Press.
Katzenstein, Lawrence C. 1989. "PRC Relations with Hong Kong: An AI Model of
Decision Making in Turbo Prolog." Northeastern Political Science Association,
Philadelphia.
Kaw, Marita. 1989. "Predicting Soviet Military Intervention." Journal of Conflict

Resolution 33,3: 402-429.
Klahr, Philip, and Donald A. Waterman. 1986. Expert Systems: Techniques, Tools
and Applications. Reading, MA: Addison-Wesley.
Krasner, Stephen D., ed. 1983. International Regimes. lthaca: Cornell University Press.
Laird, John E., Paul S. Rosenbloom, and Alien Newell. 1986. llniversal Subgoaling
and Chunking: The Automatic Generation and Learning of Goal Hierarchies.
Norwell, MA: Kluwer.
Lane, Ruth. 1986. "Artificial Intelligence and the Political Construction of Reality:
The Case of James E. Carter." American Political Science Association, Washington.
Leithauser, Brad. 1987. "The Space of One Breath." The New Yorker, 9 March
1987, pp. 41-73.
Mallery, John C. 1988. "Thinking About Foreign Policy: Finding an Appropriate Role
for Artificially Intelligent Computers." International Studies Association, St. Louis.
Margolis, Howard. 1987. Patterns, Thinking and Cognition: A Theory of Judgement.
Chicago: University of Chicago Press.
McCorduck, Pamela. 1979. Machines Who Think. San Francisco: Freeman.
Minsky, Marvin. 1986. The Society of Mind. New York: Simon and Schuster.
Neustadt, Richard E., and Ernest R. May. 1986. Thinking in Time: The llses of
History for Decision Makers. New York: Free Press.
Newel!, Alien, and Herbert Simon. 1972. Human Problem Solving. Englewood Cliffs,
NJ: Prentice-Hall.
Ordeshook, Peter C., ed. 1989. Models of Strategic Choice. Ann Arbor: University
of Michigan Press.
Pollack, Andrew. 1988. "Setbacks for Artificial Intelligence." New York Times, 4
March 1988.
Sanger, David E. 1985. "Smarter Machines Get Smarter." New York Times, 15
December 1985.
Schank, Roger C. 1987. "What is AI, Anyway?" AI Magazine 8: 5965.
Schank, Roger C., and Christopher K. Riesback. 1981. Inside Computer llnderstanding.
Hillsdale, NJ: Erlbaum Associates.
Simon, Herbert A. 1979. Models of Thought. New Haven: Yale University Press.
_ _ . 1982. The Sciences of the Artificial. 2nd ed. Cambridge: MIT Press.
___ . 1985. "Human Nature in Politics: The Dialogue of Psychology with Political
Science." American Political Science Review 79,2: 293-304.
Snyder, Glenn H., and Paul Diesing. 1977. Conflict Among Nations. Princeton: Princeton
University Press.
Sylvan, Donald A., and Steve Chan. 1984. Foreign Policy Decision Making: Perception,
Cognition and Artificial Intelligence. New York: Praeger.
Tanaka, Akihiko. 1984. "China, China Watching and CHINA-WATCHER," pp. 310-
344. In Donald A. Sylvan and Steve Chan, Foreign Policy Decision Making:
Perception, Cognition, and Artificial Intelligence. New York: Praeger.
Thorson, Stuart, and Donald A. Sylvan. 1982. "Counterfactuals and the Cuban Missile
Crisis." International Studies Quarterly 26: 537-71.
Waldrop, M. Mitchell. 1984. "Artificial Intelligence (1): Into the World." Science 233:
802-805 (24 February 1984).
_ _ . 1988. "Toward a Unified Theory of Cognition." Science 241: 27-29 (1 July
1988).
Weizenbaum, Joseph. 1976. Computer Power and Human Reason. San Francisco:
Freeman.
PART ONE
Conceptual Issues
and Practical Concerns
2
Artificial Intelligence and Intuitive Foreign Policy
Decision-Makers Viewed as Limited Information
Processors: Some Conceptual Issues and
Practical Concerns for the Future
He/en E. Purkitt
The application of artificial intelligence techniques and computational

models to study intentional activity in international relations reflects a diversity
of concepts, research foci, and methods. Underlying this diversity, however,
are several important commonalities. One of these commonalities involves
a shared recognition of the need to build models that are both computationally
tractable and cognitively plausible. Thus, most computational modeling efforts
seek to build upon what is already known about the role of cognitive
processes of political choosers operating as participants within some larger
context (i.e., small group, organizational, social, or cultural environment).
Although many of these symbolic models do not seek to replicate the actual
foreign policy decision-making process, most AI modelers claim that their
underlying model is consistent with what is known about the cognitive
processes of foreign policy decision-makers and of intuitive problem-solvers
in genera1. 1
Despite recent interest among computational modelers in developing
techniques capable of representing the complex and dynamic social, cultural,
and linguistic bases of political cognition of multiple actors at various levels
of analysis (see, for example, Alker, 1984; Bennett, 1989; Mallery, Hurwitz,
and Duffy, 1987; and Alker et al., Banerjee, Bennett and Thorson, Boynton,
and Sylvan, Majeski, and Milliken, all this volume), most AI modeling efforts
This research was supported by a grant from the U.S. Naval Academy Research Council. I
would also like to thank James W. Dyson, my collaborator on a longer-term research project.
Many of the ideas developed in this chapter reflect his prior contributions. However, any errors
in this manuscript are mine.
35
36 He/en E. Purkitt
continue to emphasize cognitive research and concepts related to how the

content and structure of prior knowledge influences problem understanding
and choice. This emphasis is understandable, given the intellectual roots of
computational models currently in use in international relations, computer
science, and cognitive psychology. In these fields the major effort centers
on understanding and developing computational techniques that model how
individuals store, represent, and access knowledge to understand current
situations and problems.
Unfortunately, these three research traditions are difficult ones to reconcile.
This difficulty will present problems for AI researchers in the future. One
problem, for example, is at the conceptual level. Conceptual difficulties
persist in part because much of the past descriptive research on political
decision making has not focused on the cognitive basis of choice except
to suggest that people do not follow the dictates of rational choice models
of decision making (Kegley, 1987; Powell, Purkitt, and Dyson, 1987; Schrodt,
1985). Thus, the common recognition of the shortcomings of the rational
choice model evident among AI researchers complicates rather than simplifies
AI modeling. (Once choice is described as rooted in more than one motive,
modeling the cognitive process becomes considerably more difficult). Although
binary or multimotive choice models may be more flexible, and perhaps
more accurate as well, how people process information to satisfy two or
more values is difficult to model fully.
In sum, computational modelers using AI techniques have an interest,
which they share with descriptive researchers focusing on how people make
decisions under conditions of uncertainty, in constructing a new process-
based paradigm of foreign policy and international relations. 2 However,
dissatisfaction with past research that has either taken a black-box approach
to the foreign policy process or used "as if" reasoning to establish the
plausibility of a rational choice paradigm, does not facilitate model devel-
opment. How, for example, Simon's (1958, 1959, 1976) satisficer thinks
through a problem is highly varied and uncertain-one simply decides when
one subjectively believes enough is known to make a choice. Thus, subjective
decisions may vary across situations and choosers.
Although it is not possible to offer concrete theoretical solutions to many
of the problems AI models in international relations must address, three
major insights that have emerged over the past several decades on how
people actually process information to solve problems intuitively seem to
offer some useful heuristic guidelines for future AI modeling efforts in
international relations. These information-processing insights are important
as they underscore the fact that there are patterned regularities in the way
people structure and process incoming information to solve problems in-
tuitively. Thus, cognitive variety in political problem solving may not be as
open-ended as the satisficing perspective implies.
AI and Intuitive Foreign Policy Decision-Makers 37
There are several reasons for relying more on information-processing

insights in future AI modeling efforts. One of the most important reasons
is that past research on how people process information intuitively under
conditions of uncertainty underscores the critical role of incoming information
and other external stimulus cues as determinants of choice in a given
situation. These results suggest there may be problems with models that
rely primarily on a priori assumptions about "the primary mechanisms"
used by intuitive processors. Given the primitive nature of our understanding
of the determinants of human cognition, it may be more productive to use
inductive-based approaches in constructing models to "fit" what is already
known about actual decisional processes. There may be a danger at this
early stage of becoming "locked in" on a particular type of model (i.e.,
knowledge systems based on predicate logic and justified in terms of the
pervasive use of analogical reasoning by real-world foreign policy decision-
makers), a model that does not conform with what is already known about
how people actually think about and make foreign policy decisions. 3
Greater attention to the insights about how people actually structure and
process information may also provide several heuristic functions for com-
putational modelers. A fundamental issue to be addressed in constructing
a computational model pertains to the appropriate level of analysis (i.e.,
individual, group, or aggregate). Even if this question depends on the purpose
of the analysts, information-processing insights suggest the existence of
important cross-level generalizations that can and should be incorporated
into computational models designed to simulate some aspect of human
cognition. At a minimum, greater familiarity with these information-processing
insights should help sensitize modelers to some of the more difficult conceptual
issues and practical concerns related to the role of external stimulus cues
and cognition that need to be addressed in future efforts to simulate aspects
of intentional behavior in international relations.
Basic Information-Processing lnsights

Research on how people process information has tended to highlight the
fact that political decision-makers act more like limited information processors
than rational maximizers or incremental satisficers. This information-processing
image of the "typical" political decision-makers is based on evidence that
has been accumulating over decades, on intuitive problem solving and
decision making. Research on how people solve complex problems has
tended to converge around a few basic generalizations about human cognition.
The most important of these generalizations revolve around three related
ideas: (1) People can only process a limited amount of information without
experiencing information overload. The main effect of increasing information
tends to be an increase in subjective confidence in the adequacy of one's
38 He/en E. Purkitt
chosen solution. Because decision-makers under conditions of uncertainty

routinely seek increased amounts of information, a pervasive characteristic
of intuitive decision making is an unwarranted faith in the adequacy of
intuitively based problem solutions. This faith has been termed "cognitive
conceit. "4 (2) People use heuristics or simplifying mental rules of thumbs
as cognitive aids during all stages of intuitive decision making. (3) There
are recurring patterns in the way people structure and process information
(i.e., both their logic and the type of information they stress is patterned).
When considering these three generalizations, we need to consider that the
specific components of the information-processing routines used at any
particular point in the process will be determined by both internal cues
(related to the organization and structure of prior experiences and knowledge
in long-term memory) and external cues (e.g., the amount and type of
information and the key aspects of incoming information perceived to be
relevant).
Intuitive Problem Solving

On the basis of information-processing research we may describe intuitive
problem solving as highly "adaptable." Evidence from experimental research
indicates that individuals adapt to the perceived demands of the immediate
task at hand. Thus, in a laboratory setting it is rather easy to induce
dramatic preference reversals and to demonstrate that people do not follow
the dictates of normative models of choice. Seemingly minor variations in
the amount and type of information made available to experimental subjects
or minor changes in the way an experimental task is presented can profoundly
affect all subsequent stages of intuitive problem solving (see, for example,
Carrell, 1980; Carrell and Payne, 1976; Kahneman, Slovic, and Tversky,
1982; Payne, 1980; Slovic, Fischhoff, and Lichtenstein, 1977, 1984; Slovic,
Lichtenstein, and Fischhoff, 1988). Thus, the highly adaptable nature of
intuitive decision making and the use of intuitive heuristics to understand
and to solve complex problems often lead to fundamental errors or biases
in the processing of information. The evidence to date suggests that the
overutilization of simple intuitive heuristics and the underutilization of formal
logical or statistical rules causes errors and biases at each of five decision-
making stages (Nisbett and Wilson, 1977; Nisbett and Ross, 1980, Pitz,
1977, 1980; Pitz and Sachs, 1984):
1. describing and coding information during the initial stage of problem

identification,
2. explaining important causal dimensions and making intuitive predictions
during the initial stage of problem identification,
3. explaining important causal dimensions and making intuitive predictions

during the problem-evaluation stage,
4. choosing among alternatives during the decision stage, and
5. evaluating, learning, integrating, and adapting to feedback information
on the efficacy, adequacy, and accuracy of prior behaviors and actions.
The essence of this line of research has been aptly stated by Nisbett
and Ross (1980: 12):
In ordinary social experience people often look for the wrong data, often see
the wrong data, often retain the wrong data, often weight the data improperly,
often fail to ask the correct questions of the data and often make the wrong
inferences on the basis of their understanding of the data. With so many
errors on the cognitive side, it is often redundant and unparsimonious to look
for motivational errors. We argue that many phenomena generally regarded
as motivation (for example, self-serving perceptions and attributions, ethno-
centric beliefs, and many types of human conflicts) can be understood better
as products of relatively pervasive information processing errors than of deep-
seated motivational forces.
Recognition that intuitive problem solving is highly adaptable to perceptions

of the nature of the immediate task and that human inferences are often
subject to systematic errors and biases when compared to normative models
of choice helps us to appreciate why it is difficult to model how people
will decide across a variety of political situations. To overcome this difficulty
we need to find the recurring patterns in the way people structure information
and, next, to use these patterned regularities to gain a reasonably accurate
description and perhaps also formal representations of the problem-solving
logic used by intuitive decision-makers.
Some of the requisite descriptive research has already been completed.
This research has indicated that when confronted with a new but complex
problem, intuitive problem-solvers seem to act like naive scientists; that
both novice and experienced political problem-solvers first try to gain a
crude sense of the problem and then develop a response; and that how the
problem is initially understood (framed) conditions all subsequent stages of
problem solving. 5
Initial framing seems to be accomplished by evaluating and applying
available information judged to be highly plausible, salient, and valid by
intuitive decision-makers (Nisbett, Borgida, Crandall, and Reed, 1976; Nisbett
and Ross, 1980; Pitz, 1980; Ross, 1977; Kahneman, Slovic, and Tversky,
1982), by seeking additional information (Einhorn and Hogarth, 1978, 1980,
1981, 1985; Slovic, Fischhoff, and Lichtenstein, 1977; Slovic and MacPhillamy,
1974), and by searching memory for relevant precedents, analogies, or clues
40 He/en E. Purkitt
to aid in the categorization of the current problem and to structure a

response (Mefford, this volume).
Cognitive Limitations
Generally speaking, the power and complexity of human cognition is tied
to the almost unlimited capacity of humans to store and combine vast
amounts of information in long-term or associative memory. Presently,
researchers do not fully grasp the structure of human knowledge and the
processes used to access and to modify prior knowledge. Conceptual issues
related to the representation of human memory and the mechanisms used
to access and to modify human knowledge are fundamental, unresolved
puzzles in cognitive psychology, AI, and computational modeling in inter-
national relations.
An important aspect of human thinking is that humans operate under
severe cognitive constraints because of the limited capacity of working or
short-term memory. Miller's (1956) suggestion that people can only retain
seven (plus or minus two) pieces of information in working memory has
been confirmed by memory-tracing research (Hayes, 1981; Payne, 1980;
Newell and Simon, 1972).
Research has also demonstrated that the active processing of information
is a serial process within the limited capacity of working memory (Payne,
1980: 95). In moving from the level of pieces of information to the level
of factors or indicators, it is now clear that individuals can only systematically
process information on a few (probably two or three) factors without explicit
cognitive aids (e.g., algorithms). Past information-processing research sug-
gests that the number of dimensions a person can systematically process
at any one time is extremely limited and may be no more than one and
certainly not more than five (Dyson, Godwin, and Hazlewood, 1974; Dyson
and Purkitt, 1986a; Heuer, 1978; Shepard, 1978).
The exact number of dimensions employed appears to vary and depend
more on such factors as the nature of the immediate task at hand and the
amount and type of information presented rather than on variations in the
prior knowledge (expertise) or experiences of individual choosers. The
experimental literature has repeatedly documented this pervasive tendency
of people to use a limited (one to five) number of dimensions in processing
information, and the critical role of information and task characteristics
across a variety of problems and task settings.
We can perhaps illustrate this tendency and the influence of task char-
acteristics and information with the findings from one line of experimental
research on how people's perceptions of risk vary with different types of
information and task settings. Slovic and Lichtenstein (1968) and Payne
(1975) found that in a gambling experiment that presented risk in terms
of four dimensions, subjects tended to perceive the probability of losing as

the most important dimension. Slovic, Fischhoff, and Lichtenstein (1983)
and Johnson and Tversky (1983) in a series of related risk experiments
found that two or three factors accounted for over 80 percent of the variance
in the dimensions of risks used by subjects.
Collectively, these and other related experiments indicate that the structure
of decision making across specific problem domains evidence a remarkable
degree of stability in terms of the number of analytical dimensions used.
People generally do not seem to be able to analyze a large amount of
information on the basis of complex multidimensional intuitive analysis.
Instead, we find that as information increases, people appear to shift
dimensions, discarding earlier information, even initially valued analytical
dimensions, without being conscious that they are in fact employing a
switching strategy during analysis. 6
A number of theoretical frameworks have been offered as fundamental
heuristics for studying foreign policy decision making. Two of the more
prevalent ones are Holsti's (1962) belief system approach and George's (1969;
1980) operational code framework. Researchers using these perspectives
often assumed a highly complex underlying cognitive structure. Although
the results from empirical research using these a priori models of cognition
confirm that individuals do use prior beliefs, operational codes, and schema,
most studies suggest that the structure of these cognitions is rather simple.
Holsti's (1962) study of John Foster Dulles concluded that a few belief
dimensions were adequate for describing Dulles's "image" of the Soviets
over time. Empirical research using George's operational code framework
rarely focuses on more than four and often as few as one element of the
code to understand changes in U.S. foreign policy (Lampton, 1973; Rosati,
1984; Walker, 1977). Thus, the fundamental insight that people only use a
few dimensions of a problem in forging solutions has filtered into the literature
on foreign policy decision making.
The limitation on the numbers of dimensions used in decision making
may operate at higher levels of organized cognitive activity. For example,
Sylvan, Majeski, and Milliken (this volume) note in reviewing the record
concerning various phases of the Vietnam ground war during the 1964-
1965 period that as a general rule, at any given time, the president and
his top advisers simply did not consider more than three principal policy
lines. 7 Although they speculate that this "rule of three" may be tied to
either cognitive limitations or structural and functional interorganizational
distinctions, the experimental evidence to date suggests that cognitive
limitations are an important aspect of the explanation.
In a recent reexamination of U.S. decision making during the Cuban
Missile Crisis, Sylvan and Thorson (1989) found similar evidence that political
decision-makers focus on a limited number of dimensions of a problem in
42 He/en E. Purkitt
forging political solutions intuitively. In the original JFK/Cuba production

system, Thorson and Sylvan (1982) posited that a model representing
knowledge as semantic kernels would be adequate. Their model was based
on archival materials and memoir literature available in the early 1980s that
indicated that U.S. decision-makers considered six separate options (i.e.,
blockade, air strike, etc.) during the crisis. This interpretation of the decision-
making process assumed that the blockade option, for example, emerged
in a sequential fashion as a separate option. However, the more detailed
transcripts of actual ExCom proceedings now available indicates that U.S.
decision-makers did not focus on a number of options in a sequential fashion
but instead quickly limited their focus to a few alternative option packages.
Thus, the blockade option surfaced as an early follow-up action within a
large set of possible military actions. Consequently, in re-examining the
content validity of their original assumptions in light of recently released
transcripts and other heretofore classified documentation of ExCom delib-
erations, Sylvan and Thorson (1989) concluded that their original model did
not adequately fit the actual process JFK and his advisers followed in
processing information and structuring a few "packages of alternatives"
involving multiple, simultaneous actions. 8
Basically, the historical record in this particular case supports a large
amount of experimental evidence indicating that intuitive problem-solvers
usually focus on a few aspects of a problem in their quest to find a problem
solution and quickly narrow their consideration of alternative solution paths
to a limited number of "option" packages. In fact, a careful reading of the
experimental evidence on how people actually process information suggests
that intuitive political problem-solvers will usually structure their discussion
of problem solutions into a series of binary choices (Dyson and Purkitt,
1986a).
From this information-processing insight practical implications can be
derived for future efforts to construct computational models. One implication
is to underscore the difficulties involved in attempting to construct com-
putational models of decision processes from retrospective accounts (i.e.,
memoirs and secondhand accounts) of actual decisional processes. There
is usually a large gap between the accounts of the deliberation and choice
processes of small-group decision making or policy proceedings that are
based on secondary sources and retrospective accounts when compared to
the actual historical record (i.e., video tapes in laboratory settings or adequate
written records for significant small-group deliberations and choices within
an organizational setting). Apparently, individuals tend to embellish what
took place in their meetings. In view of the limited amount of past research
on the actual cognitive processes and information-processing routines used
by foreign policy decision-makers, this gap raises some serious questions
about the ability of production systems to adequately represent knowledge
in "cognitively plausible" ways in the absence of a rich record of actual

group proceedings. This observation in turn seems to highlight the need
for a greater emphasis on developing highly flexible, conceptually neutral,
inductive-based AI methodologies (see Alker et al., Boynton, and Sylvan,
Majeski, and Milliken, this volume).
The Use of Heuristics and Resulting Mental Biases

To cope with limited cognitive capabilities, individuals selectively process
information and use a limited number of heuristics or mental rules of thumb
as cognitive aids in their effort to manage information. This apparently
universal reliance on intuitive heuristics to solve all types of problems seems
to be due to the need to compensate for the limitations of short-term
memory and information-processing capabilities. By using intuitive mental
heuristics, people can develop a problem attack plan that permits them to
develop a subjectively acceptable problem solution.
Simon's description of political decision making as bounded rationality
involving a satisficing process (1958, 1959), Wildavsky's (1964) description
of the incremental chooser, and Lindblom's model of muddling (1959) all
represent efforts to describe the intuitive metaheuristics frequently relied on
by political decision-makers. An intuitively grounded heuristic used across
a wide variety of questions relies on past problem-solving experiences; a
vague sense of the knowns and unknowns of a problem (i.e., topic/issue/
question); what sort of data, conditions, and situations are relevant to the
problem; and a subjective estimate of the consequences that follow from
alternative courses of action. As a result of their use of this problem-solving
approach, political decision-makers are hindered by an unwieldy problem
attack plan that highlights a feedback loop between proposed options and
inferred consequences. Such a plan lacks an objective or systematic means
of evaluation because of its essential subjectivity. Thus, lessons of history
or learning from past experiences generally are improperly drawn (Anderson,
1981; George, 1980; May, 1973). Instead, adaptation appears to be largely
a function of trial-and-error learning done on a case-by-case basis (Mefford,
1987 and this volume).
Four findings from social psychological research on the tendency of
humans to use simplifying heuristics appear to be very relevant to At-based
approaches in international relations. First, the results of experimental
research on intuitive decision making within the highly decomposed envi-
ronment of a laboratory, which focuses on how people strive to solve highly
constrained problems within fixed time periods, suggest that the specific
heuristic used during problem understanding and subsequent stages of decision
making are heavily influenced by the amount and type of information made
available to subjects and perceptions of the nature of the task at hand (see,
44 He/en E. Purkitt
for example, Carron and Payne, 1976; Dawes, 1988; Einhorn and Hogarth,
1978, 1985; Kogan and Wallach, 1967; Slovic, Lichtenstein, and Fischhoff,
1988; Payne, 1980; Wallsten, 1980). These findings suggest that AI modelers
will need extremely rich data sources, including information about key
interaction processes and informational inputs, in order to build "cognitively
plausible" models to approximate actual foreign policy decisional processes
(see, for example, Mefford and Sylvan, Majeski, and Milliken, this volume).
Second, framing research has documented quite conclusively the subtle
power of initial information to manipulate the way people think about a
problem. How a question is framed may affect all stages of intuitive decision
making for both laypersons and experts (Slovic, Fischhoff, and Lichtenstein,
1984; Kahneman, Slovic, and Tversky, 1982). For example, framing research
indicates that people are more sensitive to negative consequences than
positive ones and are more worried about losing money than they are eager
to gain the same amounts of money. Experiments in this area (Slovic,
Lichtenstein, and Fischhoff, 1988; Tversky and Kahneman, 1983; Quattrone
and Tversky, 1983) have also found that people are more willing to accept
negative outcomes if they are considered as "costs" rather than "losses"
and tend to overvalue the complete elimination of a lesser hazard while
undervaluing the reduction of a greater hazard. These findings suggest the
importance of combining AI computational techniques with experimentation
in order to learn more about the determinants of framing, particularly in
the context of small-group decision making (see, for example, Hurwitz,
1988).
A third relevant result for At-based research concerns the experimental
evidence showing that people do not use heuristics in a consistent or
integrated fashion. Rather than integrate information in a coherent way,
people tend to shift dimensions as more information becomes available
without tying the old information to the new (Estes, 1978; Einhorn and
Hogarth, 1978, 1981; Pitz, 1977). Thus, social psychological experimental
evidence on the use of heuristics underscores the crucial role played by
information cues and situational factors in addition to the structure of prior
experiences, beliefs, and knowledge bases for shaping the initial problem
understanding and thereby subsequent decision making. This line of research
suggests the importance of developing computational methods capable of
modeling inductive inference processes, trial-and-error learning both within
the context of small groups and across actors at different levels of abstraction
(see Mallery, 1987, 1988; Schrodt, this volume). 9
Fourth, computational modelers might also consider that a variety of
perspectives have been proposed to explain the pervasive tendency of people
to use simple and often widely shared perspectives to explain behavior both
in experiments and in the world generally: Kahneman, Slovic, and Tversky
(1982) posit that people use a few metaheuristics such as representativeness,
availability, and anchoring in all intuitive analysis tasks; attribution theorists

propose that people operate as naive scientists (Kelley, 1971); Chapman and
Chapman (1969) and others use cognitive schema theories (Abelson, 1981)
or emphasize the importance of sterotypic or culturally shared rules and
knowledge (Hamilton, 1979; Hamilton and Gifford, 1976; Taylor, Fiske, Etcoff,
and Ruderman, 1978). Although this proliferation of theories might suggest
diversity in the sources of people's intuitive attributions, they actually share
a common focus on the idea that what people are actually doing in introspective
analyses to make causal connections is to rely upon and make relatively
simple judgments about the extent to which some input is representative,
plausible, or consistent with some output.
Modeling External Cues

The specific stimulus cues used for making attributional linkages seem
to depend upon both the number and types of external stimulus cues
available and the results of individuals searches through easily recalled stored
connotative relations in memory (Nisbett, Borgida, Crandall, and Reed, 1976;
Pitz, 1977, 1980). A major task still facing AI researchers is to determine
how to model the importance of external stimulus cues.
In considering this problem computational modelers might find it useful
to review the evidence suggesting that perceptions of the level of uncertainty
associated with the immediate task at hand play a critical role in structuring
the problem in working memory. For example, Mitchell, Russo, and Pennington
(1989) discuss evidence that suggests that the amount of uncertainty
associated with outcomes rather than temporal sequence has the greatest
impact on the nature of explanations used for events. These ideas may
relate to studies reported by Tversky and associates. Tversky (1977), Tversky
and Heath (1989), and Tversky and Kahneman (1983) present evidence
indicating that people found prediction easier to make than explanation.
What this pattern may indicate is that predictive tasks (which seem like
games of chance) are less cognitively demanding than explanatory tasks
(which seem like games of skill). Thus, the perceived nature of the task
may be a confounding factor in the effort to model human thought.
In contrast to the focus on the role of external cues and perceptions of
the immediate task in research on intuitive judgment and decision making
under uncertainty, most AI research has been concerned primarily with
ways to model internal stimulus and prior knowledge. Mefford captured the
current emphasis on the central posited role that precedent-based analogical
reasoning and the use of relevant political analogies for structuring the
political thought and subsequent actions of a cognitive agent has when he
explained that
46 He/en E. Purkitt
we focus on how key actors manage to "construct realities," that is, to select
out what is critical in a situation (including evidence of threat or opportunity)
and to formulate appropriate courses of action. We argue that much of the
cognitive work involved in interpreting situations essentially entails posing and
reworkin~ historical analogies. In short, real or hypothetical situations-in our
case a crisis-is understood against the backdrop of selected past incidents
(1987: 222).
However, there is a problem with the approach Mefford outlines. Analogical

reasoning is clearly a primary cognitive mechanism, but AI researchers
encounter difficulties capturing the role of analogies within simple a priori
coding categories. As Sylvan, Majeski, and Milliken note, analogies "do not
come prepackaged with a whole series of features that can be plucked like
arrows from a quiver. . . . Instead, we find that analogical features are
adduced in the process of argumentation" (this volume). Their finding that
high-level bureaucrats involved in U.S. decision making about Vietnam rarely
produced more than one or two analogies seems to "fit" with a general
perspective of political decision-makers as limited information processors.
The problems to date associated with the use of a priori coding schemes
seem to be moving increased numbers of computational modelers toward
more inductive-based approaches that build upon what is already known
about human information processing. In his AI research Schrodt uses insights
about limitations of working memory and the prevalence of pattern recognition
and the use of precedent-based logic in accessing long-term memory to
illustrate an important isomorphic similarity between human and machine
information processing. For Schrodt, this provides cognitive reasons why
AI methods based on pattern recognition and a precedent logic approach
will be more intuitively pleasing to potential users than axiomatic approaches
(1985, 1987, this volume; see also Mefford, this volume).
Boynton's work (this volume, 1988, 1989; Boynton and Kim, 1987) is
relevant to our thesis that by blending AI and information-processing research
more progress may be made in model development. Boynton starts with
the assumption (supported by extensive past experimental research) that
there are consistent patterns in the way senators serving on the Senate
Agricultural Committee, the Senate Armed Services Committee, the Senate
Environmental Protection Subcommittee, and the Senate Foreign Relations
Committee frame the issues facing them, the procedures they use for
assessing recommendations made to them and in the legislative provisions
used and reused as new circumstances arise (1988, 1989, this volume). By
carefully reconstructing the details of the Senate Foreign Relations Committee
members' deliberations in the form of a "reconstructed narrative," Boynton
is collecting the type of detailed data we feel is necessary to adequately
model political decision-making processes in cognitively plausible ways.
Instead of immediately building an AI model based on off-the-shelf methods

and a large number of a priori assumptions, Boynton is proceeding inductively
to develop a text-base file that includes information on the events in the
historical record that are highly salient to this particular actor (the Senate
Foreign Relations Committee), while also retaining critical contextual and
situational information. In addition, Boynton's approach relies on the type
of highly detailed interaction data that may be necessary to build a "cognitively
plausible" model of information processing and intentional inference in a
political context. Finally, Boynton's descriptive analysis seems to provide
empirical support for the information-processing proposition that political
decision-makers only use a limited number of dimensions (i.e., beliefs,
problem factors, policy options, analogies, causal mechanisms, etc.) in
developing a representation of a problem and a proposed solution intuitively.
His finding (this volume) that there were only three times during the question-
and-answer period when congressional committee members or witnesses
strung together a longer set of "interpretive triples" (plausible pairwise
connections) to establish causal explanations for unexplained events in
constructing a historical narrative and as a basis for counterfactual arguments
and predictions during committee deliberations, is both consistent with past
experimental research and an information-processing perspective of intuitive
decision making and may provide some useful practical guidance for con-
structing future AI models if it is replicated in other politically relevant
conversational systems.
Conclusion
By combining descriptive research on how intuitive decision-makers
actually decide with AI techniques on how prior knowledge and incoming
information interact, progress in more accurately modeling political decision
processes may follow. But an essential aspect of intuitive decision processes
will confound these modeling efforts, to wit: Important differences in task
environments seem to influence political decision making. As Boynton
demonstrated, the Senate Agriculture Committee had a well-developed per-
spective of agricultural policy, and they used this shared perspective in
developing policy changes, whereas the Senate Armed Services Committee
lacked such a perspective and encountered greater difficulty in formulating
problem responses (Boynton, 1988, this volume).
Recognition of the pervasive use of mental heuristics is central to At-
based research and forms an important rationale for the use of precedent-
based models in international relations; it is also an important dimension
in information-processing research. These two approaches to the study of
choosing share a number of common interests. However, if individuals are
highly task-specific in the particular heuristics they employ, as information-
48 He/en E. Purkitt
processing research suggests, then computational modelers will face increased

difficulties. Although there are recurring patterns in the structure of how
people process information (in terms of the amount, the types of information
they stress, and the logic they employ), the specific information-processing
routines applied will vary by situation and by decision-maker. Past research
on actual intuitive inference processes suggests that a number of external
cues also influence the choice process. We have tried to show in a number
of ways how this confounds_ AI modeling efforts without dooming them as
political decision-makers only process a limited number of dimensions (i.e.,
beliefs, problem factors, policy options, analogies, or solutions) of a problem
within the confines of working memory. In fact, more attention to the
patterned regularities in the ways people process information may help in
developing computational techniques capable of modeling inductive inference
processes.
Notes
1. The question whether success in developing machine intelligence is fully
dependent on understanding the substance and principles of how humans acquire
and use knowledge is a controversial subject in the general artificial intelligence
literature yet there is a remarkable consensus among artificial intelligence modelers
in international relations on the need to develop "cognitively plausible" models. This
position means that AI modelers in international relations must consider a host of
issues related to the process validity of their models. See Mallery (1988) and Thorson
and Andersen (1987) for a discussion of validity issues that must be addressed.
2. Relevant past descriptive research on decision making under conditions of
uncertainty includes early research on human problem solving (i.e., Newell and
Simon, 1972; Simon and Hayes, 1976), experimental studies in social psychology
(see for example, Carroll and Payne, 1976; Dawes, 1988; Einhorn and Hogarth, 1978;
Estes, 1978; Janis and Mann, 1977; Kahneman, Slovic, and Tversky, 1982; Pitz,
1980; Pitz and Sachs, 1984; Rohrmann, Beach, Vlek, and Watson, 1989; Slovic,
Lichtenstein, and Fischhoff, 1988), and experimental studies in political science. The
relevant experimental research in politics is reviewed in Dyson and Purkitt, 1986a,
1986b; Kirpatrick et al., 1976; see also Bennett, 1981; Davis, 1978; Dyson, Godwin,
and Hazlewood, 1974; Purkitt and Dyson, 1988; Tetlock and McGuire, 1986. Thorson
and Andersen (1987) provide a useful introduction to the relevant past descriptive
decision-making research in social psychology for artificial intelligence applications.
3. Precedential-based logic underlies most recent AI models developed in inter-
national relations (see Mallery, 1988; Mefford, this volume). Efforts to develop
computational techniques capable of representing inductive inferences and trial-and-
error learning are only now being developed. See, for example, Boynton, and Schrodt
(this volume). See also recent applications using the RELATUS system (Mallery,
1987, 1988; Mallery, Hurwitz, and Duffy, 1987).
4. For evidence of the pervasiveness of cognitive conceit and the general tendency
for increasing amounts of information to lead to increases in subjective confidence
without a concomitant increase in the quality of intuitive decision making, see

Einhorn and Hogarth, 1985; Kogan and Wallach, 1967; Heuer, 1978; and Shanteau,
1989.
5. Although there are important differences in the way prior knowledge is structured
and stored in the associative memory of novices and experts, the experimental
evidence suggests that both novices and experts are highly ineffective at intuitive
judgmental tasks under conditions of uncertainty. See Shanteau (1987, 1989) and
Purkitt and Dyson (1987) for a review of past research on how experts and novices
process information intuitively. See Anderson (1988) and Boynton (1989) for discussions
of some possible ways that findings about expert-novice differences in experimental
research may be related to our understanding of the policy process.
6. For evidence that people use a multistage process to interpret information
and employ a number of different decision rules in their efforts to solve complex
problems, see Tversky and Kahneman (1979, 1986), Montgomery and Svenson (1976),
Park (1978), Wallsten (1980), and Dahlstand and Montgomery (1984). See Gallhofer,
Saris, and Schellekens (1988) for a discussion of nonexperimental evidence of the
same sort of patterning in the deliberations of "real-world" foreign policy decision-
makers.
7. For evidence from another computer-based study of political decision making
during the Vietnam War that was based on textual materials and found relatively
simple structuring in terms of decisional activity, see Andrus and Powell (1975).
8. See Purkitt and Dyson (1989) and Purkitt (1990, forthcoming) for a detailed
analysis of the ExCom deliberations during the Cuban Missile Crisis using an
information-processing perspective. Several other analyses have also concluded that
U.S. decision making during this crisis departed markedly from the descriptions
offered by participants in their memoirs. See, for example, Anderson (1983).
9. Recent computational models developed by Alker and his associates illustrate
one potentially promising approach because the RELATUS system is a highly flexible
system that can be used to code verbal data and to model organizations as
conversational processing systems. See Alker (this volume); Bennett (1989); Mallery
(1988); Mallery, Hurwitz, and Duffy (1987); and Mefford (this volume) for some
artificial intelligence applications of modeling organizations as conversational pro-
cessing systems.
Bibliography
Abelson, Robert P. 1981. "The psychological status of the script concept," American
Psychologist, vol. 36, pp. 715-29.
Alker, Hayward R., Jr. 1984. "Historical argumentation and statistical inference:
towards t:nore appropriate logics for historical research," Historical Methods, vol.
17(3), pp. 164-173.
Anderson, Paul A. 1981. "Justification and precedents as constraints in foreign policy
decisionmaking." American Journal of Political Science, vol. 25, pp. 738-61.
- - - . 1983. "Decision making by objection and the Cuban missile crisis,"
Administrative Science Quarterly, vol. 28, pp. 201-22.
50 He/en E. Purkitt
- - - . 1988. "Novices, experts and advisers," paper presented for the annual
meeting of the International Studies Association, March 29-April 1, St. Louis,
Missouri.
Andrus, David, and Powell, Charles A. 1975. "Thematic content analysis of decision-
making processes: the Pentagon Papers," paper presented at the International
Studies Association West Annual Conference, San Francisco.
Bennett, James P. 1989. "The ABM and INF Treaties as computational rules of
superpower competition, or an essay on inhuman understanding," paper presented
at the annual meeting of the International Studies Association, 28-31 April,
London.
Bennett, W.L. 1981. "Perception and cognition: an information processing framework
for politics," in S.L. Long (ed.), Handbook of Political Behavior, vol. 1. New York:
Plenum, pp. 69-193.
Boynton, G.R. 1989. "Language and understanding in conversations and politics,"
paper presented for the annual meeting of the Midwest Political Science Association.
___ . 1988. "Micro managing in foreign policy and agricultural policy: commu-
nication and cognition in policy making," paper presented at the 1988 meeting
of the International Studies Association.
Boynton, G.R., and Kim, C.L. 1987. "Political representation as information processing
and problem solving," paper presented at the 1987 meeting of the Southern
Political Science Association.
Carron, J.S. 1980. "Analyzing decision behavior; the magicians's audience," in T.S.
Wallsten (ed.), Cognitive Processes in Choice and Decision Behavior. Hillsdale,
NJ: Lawrence Erlbaum.
Carron, J.S., and Payne, J.W. (eds.) 1976. Cognition and Social Behavior. Hillsdale,
NJ: Lawrence Erlbaum.
Chapman, L.J., and Chapman, J.P. 1969. "Illusory correlation as an obstacle to the
use of valid psychodiagnostic signs," Journal of Abnormal Psychology, vol. 74,
pp. 271-280.
Dahlstand, U., and Montgomery, H. 1984. "Information search and evaluation processes
in decision making: a computer based process tracing study," Acta Psychologica,
vol. 56, pp. 113-123.
Davis, D.F. 1978. "Search behavior of small decision making groups: an information
processing perspective," in R.T. Golembiewski (ed.), The Small Group in Political
Science. Athens, Ga: University of Georgia Press.
Dawes, Robyn. 1988. Rational Choice in an Uncertain World. San Diego, Ca.: Harcourt
Brace Jovanovich.
Dyson, James W., Godwin, H.B., and Hazlewood, L. 1974. "Group composition,
leadership orientation and decisional outcomes," Small Group Behavior, vol. 1,
pp. 114-28.
Dyson, James W., and Purkitt, Helen E. 1986a. "An experimental study of cognitive
processes and information in political problem solving," final report to the National
Science Foundation, Aorida State University and U.S. Naval Academy.
___ . 1986b. "Review of experimental small group research," in S. Long (ed.),
Political Behavior Annual, vol. 1. Boulder, Colo.: Westview, pp. 71-101.
Einhorn, H.J., and Hogarth, R.M. 1978. "Confidence in judgement: persistence of
the illusion of validity," Psychological Review, vol. 85, pp. 395-416.
___ . 1980. "Learning from experience and suboptimal rules in decision making,"
in T.S. Wallsten (ed.), Cognitive Processes in Choice and Decision Behavior.
Hillsdale, NJ: Lawrence Erlbaum.
___ . 1981. "Behavioral decision theory: processes of judgement and choice,"
Annual Review of Psychology, vol. 32, pp. 53-88.
___ . 1985. "Ambiguity and uncertainty in probabilistic inference," Psychological
Review, vol. 93, pp. 433-461.
Estes, W.K. (ed.) (1978). Handbook of Learning and Cognitive Processes, vol. 5.
Gallhofer, lrmtraud N., Saris, Willem E., and Schellekens, Maarten. 1988. "People's
recognition of political decision arguments," Acta Psychologica, vol. 68, pp. 313-
327.
George, Alexander L. 1969. "The operational code: a neglected approach to the
study of political leaders and decisionmaking," International Studies Quarterly,
vol. 13, pp. 190-222.
___ . 1980. Presidential Decision Making in Foreign Policy: The Effective Use of
Information and Advice. Boulder, Colo: Westview Press.
Hamilton, D.L. 1979. "A cognitive attributional analysis of stereotyping," in L.
Berkowitz (ed.), Advances in Experimental Social Psychology, vol. 12. New York:
Academic Press.
Hamilton, D.L., and Gifford, R.K. 1976. "Illusory correlation in interpersonal perception:
a cognitive basis of sterotypic judgments," Journal of Experimental Social
Psychology, vol. 12, pp. 392-407.
Hayes, J.R. 1981. The Complete Problem Solver. Philadelphia, Pa.: Franklin Institute
Press.
Heuer, R.J. 1978. "Do you think you need more information?" Mimeograph. October.
Holsti, Ole R. 1962. "The belief system and national images: a case study," Journal
of Conflict Resolution, vol. 4, pp. 244-252.
Hurwitz, Roger. 1988. "Reading prisoner's dilemma interactions as drama," paper
presented at the annual meeting of the International Studies Association, March
30-April 1, St. Louis.
Janis, I. L., and Mann, L. 1977. Decision Making: A Psychological Analysis of
Conflict, Choice and Commitment. New York: Free Press.
Johnson, E.J. and Tversky, Amos. 1983. "Representations of perceptions of risks."
Technical Report NR 197-058. Office of Naval Research, June.
Kahneman, Daniel, Slovic, Paul, and Tversky, Amos (eds.) 1982. Judgement Under
Uncertainty: Heuristics and Biases. Cambridge, England: Cambridge University
Press.
Kegley, Charles W., Jr. 1987. "Decision regimes and the comparative study of foreign
policy," in C.F. Hermann, C.W. Kegley, and J.N. Rosenau (eds.), New Directions
in the Study of Foreign Policy. Boston: Alien and CJnwin, pp. 247-268.
Kelley, H.H. 1971. Attribution in Social Interaction. Morristown, NJ: General Learning
Press.
Kirkpatrick, S.A., Davis, D.F., and Robertson, R.O. 1976. "The process of political
decisionmaking in groups: search behavior and choice shifts," American Behavioral
Scientist, vol. 20, pp. 33-64.
52 He/en E. Purkitt
Kogan, N., and Wallach, M.A. 1967. "Risk taking as a function of the situation, the
person, and the group," in G. Mandler, P. Mussen, N. Kogan, and M. Wallach
(eds.), New Directions in Psychology, vol. 3. New York: Holt, Rinehart, and Winston.
Lampton, D.M. 1973. "The US image of Peking in three international crises," Western
Political Quarterly, vol. 26, pp. 28-50.
Lindblom, Charles E. 1959. "The science of muddling through," Public Administration
Review, vol. 19, pp. 79-88.
Mallery, John C. 1987. "Computing strategic language: natural language models of
belief and intention," paper presented at the annual meeting of the International
Studies Association, April.
- - . 1988. "Thinking about foreign policy: finding an appropriate role for artificially
intelligent computers," paper presented at the annual meeting of the International
Studies Association, March 29-April 1, St. Louis.
Mallery, John C., Hurwitz, Roger, and Duffy, G. 1987. "Hermeneutics," Encyclopedia
of Artificial Intelligence. New York: John Wiley.
May, E. 1973. Lessons of the Past: The Use and Misuse of Power in American
Foreign Policy. New York: Oxford University Press.
Mefford, Dwain. 1987. "Analogical reasoning and the definition of the situation: back
to Snyder for concepts and forward to artificial intelligence for method," in C.F.
Hermann, C.W. Kegley, and J.N. Rosenau (eds.), New Directions in the Study of
Foreign Policy. Boston: Alien and Unwin, pp. 221-224.
Miller, George A. 1956. "The magical number seven plus or minus two: some limits
on our capacity for processing information," Psychological Review, vol. 63, pp.
81-97.
Mitchell, D.J., Russo, J.E., and Pennington, N. 1989. "Back to the future: temporal
perspective in the explanation of events," Journal of Behavioral Decision Making,
vol. 2 (1, January-March), pp. 25-38.
Montgomery, H., and Svenson, 0. 1976. "On decision rules and information processing
strategies for choices among multiattribute alternatives," Scandinavian Journal
of Psychology, vol. 17, pp. 273-291.
Newell, Alien, and Simon, Herbert. 1972. Human Problem Solving. Englewood Cliffs,
NJ: Prentice-Hall.
Nisbett, R.E., Borgida, E., Crandall, R., and Reed, H. 1976. "Popular induction:
information is not necessarily informative," in J.S. Carroll and J.W. Paynes, (eds.),
Cognition and Social Behavior. Hillsdale, NJ: Lawrence Erlbaum.
Nisbett, R.E., and Ross, L. 1980. Human Inference: Strategies and Shortcomings in
Social Judgement. New York: John Wiley.
Nisbett, R.E., and Wilson, P.P. 1977. "Telling more than we can know: verbal reports
on mental processes," Psychological Review, vol. 84, pp. 231-59.
Park, C.W. 1978. "A seven point scale and a decision maker's simplifying choice
strategy: an operationalized satisficing-plus model," Organizational Behavior and
Human Performance, vol. 21, pp. 252-272.
Payne, J.W. 1975. "Relation of perceived risk to preferences among gambles," Journal
of Experimental Psychology: Human Perception and Performance, vol. 21, pp.
286-94.
___ . 1980. "Information processing theory: some concepts and methods applied
to decision research," in T.S. Walls ten (ed.), Cognitive Processes in Choice and
Decision Behavior. Hillsdale, NJ: Lawrence Erlbaum, pp. 95-115.
Pitz, G.F. 1977. "Decision making and cognition," in H. Jungerman and G. DeZeeuw
(eds.), Decision Making and Change in Human Affairs. Dordrecht, the Netherlands:
Reidel.
___ . 1980. "The very guides of life: the use of probabilistic information for
making decisions," in T.S. Wallsten (ed.), Cognitive Processes in Choice and
Decision Behavior. Hillsdale, NJ: Lawrence Erlbaum, pp. 77-94.
Pitz, G.F., and Sachs, N.J. 1984. "Judgment and decision making: theory and
applications," Annual Review of Psychology. vol. 35.
Powell, Charles A., Purkitt, Helen E., and Dyson, James W. 1987. "Opening the
black box: cognitive processing and optimal choice in foreign policy decision
making," in C.F. Hermann, C.W. Kegley, and J.N. Rosenau (eds.), New Directions
in the Study of Foreign Policy. Boston: Alien and Unwin, pp. 203-220.
Purkitt, Helen E. (1990, forthcoming) "Political decision making in the context of
small groups: the Cuban missile crisis revisited-one more time," in E. Singer
and M. Hudson (eds.), Political Psychology and Foreign Policy. Boulder, Colo.:
Westview.
Purkitt, Helen E., and Dyson, James W. 1987. "Does experience and more information
lead to high quality political decisions? an information processing perspective."
Mimeograph. Annapoli~i, Md.: U.S. Naval Academy.
- - - . 1988. "An experimental study of cognitive processes and information in
political problem solving," Acta Psychologica, vol. 68, pp. 329-342. ;
___ . 1989. "Foreign policy decision making under varying situational constraints:
an information processing perspective," paper presented at the Twelfth Research
Conference on Subjective Probability, Utility and Decision Making, August 21-
25, Moscow.
Quattrone, G.A., and Tversky, A. 1983. "Causal versus diagnostic contingencies: on
selfdeception and on the voter's illusion." Technical Report NR 197-058. Office
of Naval Research, June.
Rohrmann, Bernd, Beach, Lee R., Vlek, Charles, and Watson, Stephen R. (eds.) 1989.
Advances in Decision Research. Amsterdam: North-Holland.
Rosati, J.A. 1984. "The impact of beliefs on behavior: the foreign policy of the
Carter administration," in D.A. Sylvan and S. Chan (eds.), Foreign Policy Decision
Making. New York: Praeger, pp. 158-91.
Schrodt, Philip A. 1985. "Adaptive precedent based logic and rational choice: a
comparison of two approaches to the modeling of international behavior," in U.
Luterbacher and M.D. Ward (eds.), Dynamic Models of International Conflict.
Boulder, CO: Lynne Rienner, pp. 373-400.
- - - . 1987. "Pattern matching, set prediction and foreign policy analysis," in
S.J. Cimbala (ed.), Artificial Intelligence and National Security. Lexington, Ma.:
Lexington Books, pp. 89-107.
Shanteau, James. 1987. "Psychological characteristics of expert decision makers,"
in J. Mumpower, 0. Renn, L.D. Phillips, and V.R.R. Uppuluri (eds.), Expert
Judgment and Expert Systems. Berlin: Springer Verlag.
54 He/en E. Purkitt
- - - . 1989. "Psychological characteristics and strategies of expert decision

makers," in Bernd Rohrmann, Lee R. Beach, Charles Vlek, and Stephen Watson
(eds.), Advances in Decision Research. Amsterdam: North-Holland, pp. 203-215.
Shepard, R.N. 1978. "On subjectively optimum selections among multi-attribute
alternatives," in W.K. Estes (ed.), Handbook of Learning and Cognitive Processes,
vol. 5. Hillsdale, NJ: Lawrence Erlbaum.
Simon, H.A. 1945, rev. 3d ed. 1976. Administrative Behavior. New York: Free Press.
_ _ . 1958. Models of Man. New York: John Wiley.
- - - . 1959. "Theories of decision-making in economic and behavioral science,"
American Economic Review, vol. 49, pp. 253-283.
Simon, H.A., and Hayes, John R. 1976. "Understanding complex task instructions,"
in David Klard (ed.), Cognition and Instruction. Hillsdale, NJ: Lawrence Erlbaum,
pp. 269-286.
Slovic, P., Fischhoff, B., and Lichtenstein, S. 1977. "Behavioral decision theories,"
Annual Review of Psychology, vol. 28, pp. 1-39.
___ . 1983. "Characterizing perceived risk," in R.W. Kates and C. Hohenemser
(eds.), Technological Hazard Management. Cambridge, MA: Gunn and Hain.
___ . 1984. "Behavioral decision theory perspectives in risk and safety," Acta
Psychologica, vol. 56, pp. 183-203.
Slovic, P., and Lichtenstein, S. 1968. "The relative importance of probabilities and
payoffs in risk taking," Journal of Experimental Psychology, monograph supple-
ment, vol. 78(3), part 2.
Slovic, P., Lichtenstein, S., and Fischhoff, B. 1988. "Decision making," in R.C.
Atkinson, R.J. Herrnstein, G. Lindzey, and R.D. Luce (eds.), Steven's Handbook
of Experimental Psychology, 2d ed. New York: John Wiley.
Slovic P., and MacPhillamy, D.J. 1974. "Dimensional commensurability and cue
utilization in comparative judgment," Organizational Behavior and Human Per-
formance, vol. 11, pp. 172-194.
Sylvan, Donald A., and Thorson, Stuart J. 1989. "Looking back at JFK/CBA:
considering new information and alternative technology," paper presented at the
13th annual meeting of the International Studies Association, March 27-April 2,
London.
Taylor, S.E., Fiske, S.T., Etcoff, N.L., and Ruderman, A.J. 1978. "Categorical and
contextual bases of person memory and stereotyping," Journal of Personality
and Social Psychology, vol. 36, pp. 778-793.
Tetlock, Philip E., and C. McGurie, Jr. 1986. "Cognitive perspectives on foreign
policy," inS. Long (ed.), Political Behavior Annual, vol. 1. Boulder, Colo.: Westview,
pp. 147-179.
Thorson, Stuart, and Andersen, Kristi. 1987. "Computational models, expert systems,
and foreign policy," in S.J. Cimbala (ed.), Artificial Intelligence and National
Security. Lexington, Ma.: Lexington Books, pp. 147-158.
Thorson, Stuart, and Sylvan, Donald. 1982. "Counterfactuals and the Cuban missile
crisis," International Studies Quarterly, vol. 26, pp. 537-71.
Tversky, Amos. 1977. "Features of similarity," Psychological Review, vol. 84, pp.
327-352.
Tversky, Amos, and Heath, C. 1989. "Ambiguity and confidence in choice under
uncertainty," paper presented at the Twelfth Research Conference on Subjective
Probability, Utility and Decision Making, August 21-25, Moscow.
Tversky, Amos, and Kahneman, D. 1983. "Extensional vs. intuitive reasoning, the
conjunction fallacy in probability judgment," Psychological Review, vol. 90, pp.
293-315.
___ . 1986. "Rational choice and the framing of decisions," in R.M. Hogarth and
M.W. Reder (eds.), Rational Choice: The Contrast Between Economics and Psy-
chology. Chicago, 11.: University of Chicago Press, pp. 67-94.
Walker, Stephen G. 1977. "The interface between beliefs and behavior: Henry
Kissinger's operational code and the Vietnam war," Journal of Conflict Resolution,
vol. 21, pp. 129-168.
Wallsten, T.S. (ed.) 1980. Cognitive Processes in Choice and Decision Behavior.
Wildavsky, A. 1964. The Politics of the Budgetary Process. Boston: Little, Brown.
3
Steps Toward Artificial Intelligence:
Rule-Based, Case-Based, and
Explanation-Based Models of Politics
Dwain Me/ford
New subfields within artificial intelligence emerge on a regular basis about

every five to ten years. The most recent include Case-Based Reasoning
(Kolodner 1988; Hammond 1989a, 1989b) and Explanation Based Learning
(DeJong 1988; DeJong and Mooney 1986; Mitchell, Keller, and Kedar-Cabelli
1986). These approaches cast reasoning in terms of the acquisition and use
of large qualitative structures in memory and show signs of merging into
a general approach to problem solving and planning. 1 In this chapter we
argue, as we have argued elsewhere (Mefford, 1988h, 1989), that these
approaches speak to some of the most fundamental questions in the study
of politics. The concepts and mechanisms of these approaches can be
readily applied to address paradigmatic issues in the areas of decision
making, collective action, and the theory of regimes and institutions.
The development of case-based reasoning (CBR) and explanation-based
learning (EBL) within artificial intelligence in the 1980s was motivated by
several sets of issues, including that of the organization of dynamic memory
(Schank 1982; Kolodner 1980, 1983, 1984) and the desire to overcome the
limitations of rule-based systems-limitations that have become all too
apparent after twenty years of experience. Political scientists who have built
rule-based systems have themselves become increasingly aware of these
limitations. My purpose is not to argue the virtue of one class of systems
over another, but to trace out the substantive and theoretical issues that
have motivated the development of alternative approaches. There is a
conceptual path that leads from rule-based systems, with their uniform and
atomistic data structures, to systems that use larger and more complex
constructs to reproduce inferential and learning processes. The assumption
of a fixed and static knowledge base is increasingly giving way to a fundamental
concern with how knowledge is acquired and organized dynamically.
56
Rule-Based, Case-Based, and Explanation-Based Models 57
This progression, if extended, can be steered in the direction of certain

core questions in empirical political theory. These issues include, for example,
the conceptual challenge of systematically investigating what March and
Olsen call "path dependent processes" (1984). A path-dependent process
is one in which events impact on the structure of a system, changing its
character and thereby opening and closing possibilities for subsequent change
or development (David 1985). March and Olsen express this notion cryptically
in the context of organization theory with the idea that institutions encode
their histories in their structure (March and Olsen 1984: 743). Path-dependent
processes are ubiquitous: The sexual abuse of a child may be expressed
through a lifetime of false intimacy and failed relationships; it matters for
the character of today's regimes in Eastern Europe that the German advance
was stopped at Stalingrad fifty years ago. Within the study of foreign policy
and international relations, the role of history in shaping strategic relations
and institutions is a path-dependent process, as is the impact of contingent
experience on the political beliefs of individuals. These issues, and a host
of others that are closely linked to them, can only be addressed awkwardly,
if at all, by behavioral methods (Aiker 1975, 1984; March and Olsen 1984:
749). The concepts and mechanisms that are emerging in the most recent
work in artificial intelligence provide a basis for the systematic formal and
empirical investigation of processes of this order.
To describe what is essential and characteristic of each of several
approaches, and to do so such that they can be systematically compared,
we employ a single working example that will be cast and recast in several
forms. The example is based on a classic in the literature on the violent
transfer of political power, Luttwak's study of coups d'etat (Luttwak 1979).
The patterns Luttwak finds in the planning and execution of coups will be
presented, first, as a rule-based expert system, then as a planning system
that packages and reapplies pieces of plans, then as a case-based reasoning
system that reasons on the basis of historical examples, and finally as a
system that constructs and updates its theory of coups d'etat as a function
not only of the cases that history presents but as a function of the order
in which these cases are encountered. This last class of systems paves the
way for the systematic study of the experiential learning and other path-
dependent processes that account for institutional change (Haas 1982).
The Rational Science of Artiftciallntelligence

It will become apparent as we step from system to system that the
object of the model shifts progressively from behavior, viewed in a sense
from the "outside," to the reasoning and adaptation that underlie behavior.
This effort to systematically investigate behavior from the "inside" flies in
the face of a deep prejudice of behavioral science, which, as Simon observes,
58 Dwain Me/ford
is "always wary of intervening variables that are not directly observable"

(Simon 1976: 261). But this is precisely the purpose of artificial intelligence
as it is of cognitive science in general: The object is to systematically
reconstruct the complex process that stands between the stimulus and the
response in the S-R model. In short, the theoretical action is in the mechanism
that relates the S to the R; the theoretical action is in the hyphen. Artificial
intelligence, and cognitive science more broadly (Newel! 1983a; Simon 1980),
departs from the norms of behavioral science not only in its method and
its mathematics but more fundamentally in how it defines its scientific
problem. Because conventional methodology restricts the complexity of the
problems we can study, Simon observes that
if we are to study process, we are going to have to employ research methods

and research designs that are appropriate to that kind of investigation, and
these are going to be different from the methods and designs we have used
to study simple stimulus-response connections. The shift from S-R formulations
to theory of information-processing formulations is a fundamental shift in
paradigm, and it is bringing with it fundamental shifts in method also (Simon
1976: 261).
In terms of our working example, rather than attempting to find a

relationship between, say, types of regimes and the incidence of coups d'etat
that can be expressed in some linear form, (e.g., Thompson 1974, 1975),
the object is to reconstruct, abstractly, the planning process that might be
typical of a class of powerful but disaffected political figures. Intentional
behavior is explained in terms of the capacity of intelligent agents to reason,
communicate, and act on the basis of and within the constraints imposed
by their knowledge and goals.
In the search for mechanism, artificial intelligence and cognitive science
pick up where descriptive behavioral science too often leaves off. Explanation
within this conception of science corresponds to the commonsense notion:
The object is to answer the question of how and why something happened,
or could happen. There is a stark difference between this meaning of
explanation and "explanation" as used in the notion of "explaining the
variance" between a pair of variables. 2 To use the example that we will
return to throughout this chapter, to know that military rank and age
"explain," "predict," or are "significantly" related to the likelihood that an
individual will participate in a coup is not to know why this might be the
case.
As a method of inquiry, artificial intelligence is sometimes presented as
a radical departure from the types of formal and empirical investigation
that are familiar in political science and economics (Schrodt, 1985). There
are differences in the substantive issues raised and in the function that
computational methods play in the formulation of theory, but it can be

argued that these differences remain largely on the surface. The core scientific
enterprise across the full spectrum of the AI community, from "left-wing
scruffies" to "right-wing neats" (Abelson 1981), conforms to a basic image
of science that is broadly shared. The object is to extract general principles
or mechanisms that underlie intelligent behavior.
To date, the political scientists who have applied concepts and techniques
from AI to politics have concentrated their attention on two subfields within
artificial intelligence: natural language processing, and what we will loosely
call problem solving and planning. Duffy and Mallery's project, which is
reported elsewhere in this volume, is far and away the preeminent example
of the work by political scientists in the area of natural language understanding
(Duffy and Mallery 1986; Duffy 1988; Mallery 1987, 1988a, 1988b). This
chapter concentrates exclusively on the second of the two agendas, that
of problem solving and planning, which, when viewed as a repeated process
generalizes to the question of learning. We will first chart the evolution of
AI work in planning, problem solving as it has evolved in the 1970s and
1980s. Viewed from the altitude that we adopt, the overlapping developments
in this body of work reveal a basic progression from rule-based, performance-
oriented systems to more cognitively oriented programs that utilize larger,
qualitative structures and that, increasingly, emphasize the acquisition and
modification of knowledge. The actual evolution is not, of course, as neat
and linear as our examples will suggest. (For a more complete survey that
charts the parallel lines of development and backtracking that has occurred
as these ideas have unfolded see Mefford forthcoming, chapters 3 and 4). 3
Three Generations of Rule-Based Systems

In 1967 Edward Luttwak published Coup d'Etat: A Practical Handbook,
which is remarkable as much for its format as for its findings. The book
distills the concept of coup d'etat from a nearly exhaustive data base from
the period 1945 through the early 1960s, later extended to 1978. Luttwak
explicates the concept of coup d'etat by adopting the perspective of a
hypothetical group bent on overthrowing a seated government. The general
scheme is the precipitate of 282 cases (Table 11, pp. 195-207). From these
cases, particularly from cases that strike Luttwak as paradigmatic, such as
Nasser's rise to power in 1952, he assembles a dynamic concept of coup
in terms of the opportunities and challenges that confront any would-be
organizer of an armed revolt. The book is written in the genre of the "advice
books" to princes, of which Machiavelli's is the most noted example (Gilbert
1965). It is a how-to book that prescribes the steps in the conduct of a
coup from the initial recruitment of coconspirators, through the planning
and execution, to the consolidation of power and the return to "normalcy"
60 Dwain Mefford
under the rule of a new leadership. In the preface to the second edition
(1979), Luttwak notes, with more satisfaction than chagrin, that the book
was apparently used as a guide on at least one occasion: A heavily annotated
copy of the French edition was found among the personal effects of an
ambitious colonel in an unnamed country.
Our first attempt to model the political reasoning of Luttwak's coup-
makers takes the form of rule-based system. The object is to reconstruct
the plans and calculations of a hypothetical individual or group as it organizes
and executes a coup d'etat. After observing what can and cannot be
represented in this class of systems, we proceed to other types of programs
based on principles that promise to better represent the cognitive mechanisms
at work. In addition to reconstructing the strategic reasoning of the actors,
the attempt will be to capture more fundamental processes, such as how
individuals model themselves and their political programs after historical
examples. The research question shifts, in effect, from reconstructing patterns
in data to modeling the political reasoning and political education of intelligent,
purposive agents. When we shift the object of the model in this way, Luttwak
himself, as a stand-in for an intelligent and historically informed political
actor, becomes the subject. The point is reproduce how Luttwak reasons,
how he recognizes patterns in the historical examples, and how he formulates
those patterns into principles for a particular kind of political action, namely,
armed rebellion. A theory of how regimes are overthrown requires an explicit
treatment of how individuals come to see this course of action not only as
an option but as a political goal.
Production Systems: The Cognitive and Computer Architecture

of First-Generation Rule-Based Systems
Rule-based systems and production systems, if they are not the same
concept, are so closely related that they can be safely treated as synonymous. 4
They define an architecture or principle of design for a class of computer
programs. Production systems consist essentially of a collection of rules or
productions plus a pattern matcher, or deductive procedure, or other device
for applying the rules to the problem or task. The rules or productions are
condition-action pairs, or if . . . then clauses. The condition is matched
against data or statements in the workspace. If the match succeeds, i.e.,
if a rule's condition is satisfied, then the associated action is executed,
which generally results in a change in the contents of the workspace, e.g.,
hypotheses are added or removed, or likelihoods are increased or decremented.
This conception of the function and design of computer systems is
credited to Alien Newell, building on the work of Post, Floyd, Simon, and
Feigenbaum (Post 1943; Minsky 1967; Floyd 1961; Feigenbaum 1963; Simon
and Feigenbaum 1964). Production systems have proven to be extremely
powerful and plastic. Newell, Anderson, and others have argued, and to
some extent have demonstrated, that the components of a production system
are sufficient to reproduce a wide range of intelligent behavior. The conception
is, therefore, a candidate theory of what is necessary and sufficient for
intelligence in human beings and machines (Anderson 1976, 1983; Newell
1980, 1989).5
Edward Feigenbaum is responsible for first applying the concept of
production system outside the domain of the theoretical and empirical study
of human memory (Buchanan and Shortliffe 1984: 7-8). He convinced the
designers of DENDRAL, a program designed to identify organic compounds
from mass spectographs, to use productions in place of the conventional
programming language and architecture. The insight was that the principles
in this branch of chemistry are better conceived as a collection of rules
than as some large algorithm. The success of that system (Feigenbaum,
Buchanan, and Lederberg 1971; Lindsay, Buchanan, Feigenbaum, and Led-
erberg 1981) inspired efforts not only to employ production systems in
other domains, most notably medical diagnosis, but also to develop shells
or languages to facilitate the construction of such systems across a broad
range of applications (Feigenbaum 1977). The potential that DENDRAL,
MYCIN, and related systems demonstrated in the 1970s,6 plus the advent
of special-purpose languages and development tools, including EMYCIN (Van
Melle 1980) and OPS5 (Brownston, Farrell, Kant, and Martin 1985), and the
founding of companies dedicated to capitalizing on expert system design
(Harmon 1989) is credited with "bringing AI out of the laboratory" and into
the world of business and industry (Winston and Prendergast 1984). For
this reason artificial intelligence is often identified exclusively with rule-based
systems. 7
Sophisticated rule-based systems may consist of a number of modules,
some of which elicit data from users, monitor the state of the rule base,
or offer explanations by backtracking through the chain of rules used to
reach a conclusion (Davis, Buchanan, and Shortliffe 1977). But there are
only three essential components of a rule-based system: a set of rules, e.g.,
relationships between symptoms and diseases; a data set that, in a medical
application, might include the patient's medical history and the results of
laboratory tests; and a device for applying these rules in the process of
solving a problem or executing a task (Davis and King 1984: 21). The
classical rule-based systems were primarily designed for the task of clas-
sification or diagnosis, though a diagnostic system like MYCIN is also intended
to prescribe therapies, which is a form of planning (Chandrasekaran 1986:
27-28). Classification goes hand in hand with planning in these systems,
and the same is the case for our rule-based version of Luttwak's theory of
coups d'etat.
62 Dwain Me/ford
Model 1: Luttwak Rendered as a Rule-Based System

Luttwak's account of the steps and substeps in organizing and executing
a coup clearly has the character of a planning system. The particular plan
that it, in effect, constructs, is dependent on contingent factors, such as
the ethnic divisions within the leadership circle (Luttwak 1979: 55). The
reasoning that relates these factors to what is involved in carrying out a
coup d'etat is readily represented in a rule-based system. For example, the
essential goal for the perpetrator of a coup is to take over and control the
power of the state rather than to destroy it. To do this requires neutralizing
the instruments of the state: the armed forces, the security agencies, the
bureaucracy, etc. This in turn requires infiltrating and eo-opting crucial
personnel and/or rendering their ability to oppose the coup ineffective, which
can be achieved by interfering with the channels of communication, e.g.,
seizing radio stations and silencing government officials by arresting or
kidnapping them. In a rule-based system, these actions, and a thousand
others, can each be represented as a condition-action pair of the form:
IF CONDITION, e.g., an army garrison is stationed in the capital,

THEN ACTION, e.g., reduce the effectiveness of that garrison by
disrupting its communication system.
Luttwak does not in fact formulate discrete rules, but the rulelike character
of much of what Luttwak extracts from examples of coups d'etat can be
reexpressed in such a form. For instance, as part of his analysis of the
potential for mutiny in Portugal in the 1970s, Luttwak formulates a rule-
like principle: "The forces relevant to a coup are those whose locations
and/ or equipment enables them to intervene in its locale (usually the capital
city) within the 12-24-hour time-span which precedes the establishment of
its control over the machinery of government" (1979: 70). This principle
is fleshed out via an analysis of the tactics that are likely to be effective
in blocking military units of various descriptions from coming to the defense
of the government. In a table Luttwak matches functional descriptions of
army units with the tactic that is likely to be effective in "turning" or
neutralizing that unit. Table 3.1 is effectively an array of simple rules, which
can be read off by matching descriptions of the military unit to the
corresponding tactic.
Equipped with rules of this sort, encoded in some convenient form, the
program chains backwards from the goal, i.e., to overthrow the regime,
through the actions and conditions that would realize that goal. A sample
of the rules arranged into a tree structure is pictured in Figure 3.1, in a
later section on planning systems. It is important to observe that first-
generation rule-based systems are blind to the order or dependency among
63
Table 3.1: Infiltration Strategies (Simplified from Luttwak, 1979, table 3, p. 74)
Unit: Battalion No. 1 Battalion No.2 Battalion No.3

Technical Very simple Very complex Medium
Structure: Relies on Requires Relies on
ordinary airlift and land transport
communication sophisticated and radio
and communications links
transportation
Key men Leaders 40 technicians 5 technicians
who must
be eo-opted:
Figure 3.1: AND/OR Graph of Tasks Implicit in the Concept of Coup d'Etat
to execute a coup d'etat

I
AND
recruit neutralize consolidate
~rat~rr_s t_h_~~-:r1 :_a_~_;

_____ ________Tip_o_w_e_r-----,l
identify neutralize neutralize neutralize neutralize

isaffected military police and non-state bureaucracy
leaders command security centers
OR forces of power
I I
I
by exploiting by compromising
geographical communications
dispersement AND
political parties labor
etc. etc.
64 Dwain Me/ford
the rules. That structure is implicit in the rule base, but is itself not directly
accessible to the system. As a consequence it cannot be explicitly used in
the process of monitoring a situation or solving a problem, and surfaces
only as the system attempts to solve its problem. The "flat" and "linear"
character (Sacerdoti 1977) of early rule-based systems, i.e., the principle
of design whereby all knowledge in the system is represented in the same
form at the same level, is both a strength and weakness. Such systems
are easy to modify by adding or deleting rules because there is no need
to update complex data structures. But, by the same token, because the
system cannot explicitly access this structural knowledge, it cannot utilize
it. These issues, among others, motivated the development of the next
generation of rule-based systems.
The Step to Second-Generation Rule-Based Systems:

Functional and Contextual Knowledge Plus the Problem of Explanation
What Davis has called "first-generation" rule-based or expert systems,
most notably DENDRAL (Buchanan, Sutherland, and Feigenbaum 1969;
Lindsay, Buchanan, Feigenbaum, and Lederberg 1981) and MACSYMA (Martin
and Fateman 1971; MACSYMA group 1974), were concerned exclusively
with performance. No claims were advanced as to any correspondence
between how these systems solved problems in their domains and how
human beings accomplish similar tasks (Davis 1984: 20). What mattered
was the accuracy of the result and the efficiency with which it is obtained.
In the next generation of systems, most notably the medical diagnostic
program MYCIN (Shortliffe 1976; Buchanan and Shortliffe 1984), the inter-
action between the program and human experts became an explicit concern.
For the system to be effective in a hospital setting it was necessary to
equip it with procedures for entering data and for explaining or justifying
its reasoning in a manner consistent with conventions that people use and
expect. Expert systems like MYCIN fell well short of producing convincing
explanations, but the fact that this complex issue was given prominence
on the research agenda marks an advance in the conception and design of
rule-based systems.
For a system to interact with human beings in even the restricted
conversational mode required for data entry, it must be equipped with
knowledge of a type and form that cannot be readily represented in discrete,
rulelike pieces.
In addition to being incapable of explicitly representing connections among
the rules in the rule base, rule-based systems cannot relate individual rules
to the source or evidence from which the rules were extracted. Clancy's
work shows that "individual rules play different roles, have different kinds
of justifications, and are constructed using different rationales for the ordering
and choice of premise clauses" (Ciancy 1983: 215-216). For a rule-based

system to be intelligible to a human being, what Clancy calls the "structural,"
"strategic," and "support" knowledge that underlies the rules must be made
explicit, otherwise the user cannot evaluate the system's conclusions, nor
is it possible for the user to intervene and alter the rule base appropriately
in the event that the program needs correction. Though several correctives
were proposed (Davis 1980; Clancy 1985), what is important in the present
context is that the limitations that Clancy describes in MYCIN-Iike systems
also plague any rule-based system of the same type, 8 like the one we
sketched with the Luttwak example. However, equally applicable to any
attempt by political scientists to construct similar rule-based systems is
Clancy's evaluation of the MYCIN project:
Despite the now apparent shortcoming of MYCIN's rule formalism, we must

remember that the program was influential because it worked well. . . . [W]e
can treat the knowledge base as a reservoir of expertise, something never
before captured in quite this way, and use it to suggest better representations
(Ciancy 1983: 249).
As Davis concludes, rule-based systems at present can be characterized

with the phrases "narrow domain of expertise. Fragile behavior at the
boundaries. Limited knowledge representation language. Limited 1/0. Limited
explanation. One expert as knowledge base 'czar' "(1982: 9). What is needed,
minimally, is a system capable of a "range of behaviors associated with
expertise: solve the problem, explain the result, learn, restructure knowledge,
break rules, determine relevance, degrade gracefully" (Davis 1982: 3). What
cannot easily be captured in a collection of atomized rules, even if thousands
are encoded, is the strategic knowledge of how to use them, when to allow
variations and exceptions, and how to adapt and revise them as needed.
This demands a richer representation than that which rules provide. These
limitations, coupled to the fact that rules are hard to extract because people
are often unable to articulate their knowledge in rule form, motivates the
search for alternative approaches.
Third-Generation Rule-Based Systems:

Structural Knowledge and Task Decomposition
The next generation of expert systems, e.g., Davis's work on a computer
fault-diagnostic system (Davis 1984b), requires the ability to reason in terms
of the structure and function of objects in a domain as well as the capacity
to draw inferences on the basis of empirical associations in the manner of
earlier rule-based systems. Adding structural, functional, and behavioral
descriptions entails employing multiple representations; rules no longer enjoy
a monopoly. What we are calling "third-generation" systems introduce a
66 Dwain Me/ford
new level of abstraction that focuses on the inferential tasks rather than
on the implementation of these tasks in any particular formalisms, such as
rules or frames.
The JESSE system developed by Chandrasekaran, Gael, and Sylvan (Gael
and Chandrasekaran 1987; Gael, Chandrasekaran, and Sylvan 1987; Sylvan,
Gael, and Chandrasekaran 1988 and this volume) applies this theoretical
concern with information-processing tasks to policy formation in the area
of Japanese energy planning. This program marks a major departure from
other rule-based applications to decision making principally because of the
theoretical questions that the system poses in its architecture. Programs
like Job and Johnson's reconstruction of the Johnson administration's decision
in the Dominican Republic crisis (Job and Johnson 1986; this volume) or
Thorson and Sylvan's earlier effort to reconstruct Kennedy's options in the
Cuban Missile Crisis (Thorson and Sylvan 1982) or my own work in modeling
Soviet interventions in Eastern Europe (Mefford 1984, 1987a) essentially
serve to demonstrate that it is possible to reproduce specific complex
behavior characteristic of decision making in realistic contexts. But these
models remain case-specific, as Mallery points out (Mallery 1988c). The
principal merit of these experiments is to demonstrate the utility of a
formalism or technology to capture a class of behavior in a reproducible
form. This is an important achievement; it prepares the way for, but does
not deliver, a theoretical contribution to the study of foreign policy. What
distinguishes these efforts from case studies (George and Smoke 1989;
Lebow and Stein 1989) is that, in principle, the decision mechanisms
explicated in the design of the program can serve as building blocks for a
theory of decision making by U.S. administrations, or by governments in
general, across a range of contexts. But, even though this is the end objective
of the exercise, in existing systems the logical steps from the specific case
to the general theory are incompletely worked out at best.
In contrast, by virtue of the theoretical concerns that characterize third-
generation rule-based systems, Gael, Chandrasekaran, and Sylvan pose an
explicit theoretical question through the theory of cognitive tasks embodied
in the architecture chosen for their system. They ask in effect whether it
is possible to reproduce patterns in Japanese energy policy formation using
only two basic inferential processes, that of diagnosis/classification and that
of planning. Viewed from the perspective of the theory of design of expert
systems, the performance of the system is a measure of the analysis of
the set of essential tasks or problems embodied in the Japanese case, along
with the adequacy of the high-level language in which these tasks are
implemented (Bylander, Mittal, and Chandrasekaran 1983; Bylander and Mittal
1986). Viewed from the perspective of the theory of foreign policy, the
adequacy of the system is validation of the hypothesis that policy-making
apparatus realized in the Japanese bureaucracy is an instance of a class
of information-processing systems of which the computer program is an

example. The verisimilitude of the program's performance reflects back on
the sufficiency of the underlying theory of tasks and task composition that
is operationalized in a high-level computer language that has been developed
by Chandrasekaran and his colleagues. 9
from Rules to Cases
Systems That Plan: A Bridge Between Rule-Based

and Case-Based Systems
For the purpose of making rule-based systems more transparent and less
brittle, system designers like Davis and Clancy explored such issues as that
of providing the system with "structural knowledge," or with the capacity
to trace back and question the evidence that justifies a rule. For the
theoretically oriented social scientist, the character of these knowledge
structures is of intrinsic interest. Programs serve as a means for posing
hypotheses about their properties. One of these structures, or a family of
structures, that has been an essential concern for artificial intelligence from
the beginning is that of plans and planning. The activity of planning, by
individuals or organizations, and the construction of strategic plans predicated
on the plans of other agents, is part of the core agenda of foreign policy
analysis. The applications of artificial intelligence to foreign policy are, by
and large, planning systems, for the most part cast in a rule-based form.
To make the argument that recent developments in AI, such as case-based
reasoning, can be used to pose important theoretical questions that are
difficult to pursue in rule-based systems, it is useful to first establish the
characteristics of the planning activity that the computer model should
capture. In what has been called the "classical planning framework" in
robotics (Pednault 1987; Georgeff 1987), constructing a plan consists of
ordering a sequence of actions that, when executed, progressively change
an initial state of the world into a state that satisfies a goal. A recipe, a
blueprint, or SlOP, viewed functionally, are each examples of this conception.
The notion of a plan as a set of instructions or as a program of actions
predates the major work in ropot planning. In Plans and the Structure of
Behavior, Miller, Galanter, and Pribram redefine the familiar notion of a
"plan" such that it becomes functionally equivalent to a computer program
(1960). Indeed, they capitalize the word to emphasize this special meaning.
A "Plan" is "any hierarchical process in the organism that can control the
order in which a sequence of operations is to be performed" (Miller, Galanter,
and Pribram 1960: 16). Miller et al., go further to speculate that the
instructions that make up a plan are themselves conditioned actions-that
is, actions that can be applied only given that specified conditions are
68 Dwain Mefford
satisfied, or actions that may be repeated until some condition is achieved

(Miller, Galanter, and Pribram 1960: 38). Construed in this way, plans are
quite literally programmable instructions with an explicit control structure,
i.e., it is clear under which conditions which actions will be executed.
The notion of "public plans" connects planning and problem solving at
the level of the individual with the analogous activity at the level of
organizations and institutions as conceived by Simon (March and Simon
1958). Plans, or what March and Simon in their book call "programs" (1958:
eh. 6), can be viewed as the product, qua policy, or purposive activity of
complex organizations. Unfortunately for the social sciences, subsequent
work in the field of cognitive science and artificial intelligence has, until
comparatively recently, neglected the collective dimension. Within political
science these ideas have been pursued by Crecine (1969) and by Alker and
his students (Aiker 1981; Alker, Bennett, and Mefford 1980; Alker and
Greenberg 1977; Bennett and Alker 1977; Tanaka 1984). 10
A decade after Miller, Galanter, and Pribram, the work in planning and
problem solving had progressed to the point that working systems, i.e.,
robots, or programs that simulated robots, could be built for certain restricted
domains. The most notable of these is the STRIPS program developed at
Stanford Research International (Fikes and Nilsson, 1971; Fikes, Hart, and
Nilsson 1972; Nilsson 1980: eh. 7) and Winograd's "Blocks worlds" program,
which was his thesis at MIT (Winograd 1972). Systems like STRIPS and
SHRDLU describe states of the world as a list of statements or well-formed
formulas in first order logic. The planning problem becomes one of selecting
a list of operators from those available to the system, e.g., to "move" or
"clear" blocks, or to "open" and "pass through" doors, etc., which, if
executed sequentially, transform the world from an initial state to one
corresponding to a goal. Since Newel! and Simon's work on the logic theorist,
which proved theorems in the prepositional calculus (Newell and Simon
1963, 1972), the task of planning has been construed as a task of constructing
a proof. Literally, a plan is a proof by construction in which the theorem
to be proved corresponds to a set of statements describing a goal, and the
assumptions and axioms consist of a set of initial statements plus a set of
operations, interpreted as actions. These operations transform or rewrite
statements into statements.
Under this definition of the task, a planning system can harness the
power of algorithms developed for automated theorem proving, principally
Robinson's Resolution Principle, which revolutionized that field in the mid-
1960s (Robinson 1965, 1979; Nilsson 1980: eh. 5). The engine that powers
planning systems like STRIPS is a theorem prover modified to work efficiently
with a particular data structure, e.g., lists of formulae representing the
robot's world. The system is equipped with operators, such as "pass through
door," that, when applied, have the effect of altering the state of the world
(or at least altering the program's image of the world). Changes are effected
by adding or deleting statements from the list of statements that define the
current state of the world. In later versions of the program, and in its
successors, additional procedures are incorporated for generalizing and
repositing plans, and for constructing plans in a multipass fashion that adds
efficiencies and avoids dilemmas that plagued the original system (Sacerdoti
1974, 1977; Stefik 1981). These extensions are of great interest for what
they suggest for the study of the policy process, which is manifestly a
multilevel process realized in large organizations or governments (Durfee
1986; Durfee and Lesser 1987; Durfee, Lesser, and Corkill 1987).
Model 2: Coups d'Etat Represented as Operative Plans

A system that explicitly uses the structure of plans in its reasoning is
an intuitively appealing representation of political calculations. Using our
Luttwak example, the perpetrator of a coup can be modeled as a system
that constructs and modifies plans in the face of a changeable environment.
This can be implemented using a rule-based technology, provided that the
structure of the plan itself is accessible to the system and not left simply
implicit in the rule base. What is of interest is how the system constructs
and adapts its operative plan in the face of actions and events. The conditional
structure of the plan for a particular coup might take the general form of
an andjor graph, as would the more general conception of what is involved
in seizing power. A part of that structure might take the form of the diagram
in Figure 3.1. To each of the branches in the tree corresponds one or more
rules or productions, the choice of which will likely depend on concrete
constraints and opportunities in the particular case. One of Luttwak's examples
is the failed coup attempt by the French generals in Algeria in 1961. To
overthrow a charismatic leader like de Gaulle requires not only the active
support or tacit acceptance by the military but also the power to deny that
leader the opportunity to appeal to the rank and file in the military (Luttwak
1979: 105). These factors, and others like them, would figure among the
contingencies in a fully developed plan.
In effect, the rule base and the control structure that is implicit in the
scheduling of the rules embodies a theory of coup d'etat. A "strong" theory
of this type would assert that the perpetrators of coups, and, for that matter,
all other actors involved, reason, to some approximation, in ways that
correspond to the steps that such a rule-based system takes as it solves
the problem of executing a coup in a given context. A "weak" theory makes
no claim as to the correspondence between the functioning of the program
and the reasoning and actions of people. The assertion would simply be
that the program generates behavior that, at some level, corresponds to or
predicts the behavior of some class of political systems. lt is our view that
70 Dwain Mefford
political science should strive for explanative theory in the strong sense
rather than for mere prediction. 11 The premise that the scientific purpose
should be to capture processes that operate in the world is a prime tenet
of modern scientific realism (Boyd 1984; Harre 1986; Wendt 1987). In the
present context this means seeking out those developments within artificial
intelligence that promise to provide conceptual leverage for reconstructing
the cognitive mechanisms at work in political settings. One of the most
promising developments differs from the rule-based approach both in its
data structure and in the reasoning process it attempts to reproduce.
Case-Based Reasoning as a Refinement of

the Information-Processing Model
Case-based reasoning addresses several of the limitations inherent in rule-
based systems. It expressly raises the issue whether human beings reason
on the basis of rules using rulelike deduction. In the place of large numbers
of discrete rules, people seem to rely on much more complex structures
like those of metaphor and analogy (Vosniadou and Ortony 1989). Because
rule-based systems are poor models of cognition, the builders of such
systems face not only the problem that Davis and Clancy explore-that of
making a program transparent to the user-but the related and perhaps
more serious problem of converting knowledge from the forms in which
people use it to the form that a rule-based system can apply. This extraction
and conversion process, which Feigenbaum calls "knowledge engineering"
(1977), is not only a bottleneck in the application of this technology but
may well bound the universe of problems to which these systems can be
applied. More sophisticated interfaces may not be enough. lt may be necessary
to radically rethink the fundamental design of these systems, i.e., to explore
data structures that are not rulelike and to attempt to mimic informal types
of reasoning. Case-based reasoning can be interpreted as one such effort.
If case-based reasoning is a truer model of human reasoning, then it should
be easier to transfer knowledge from human experts to systems designed
on these principles. In addition, unlike rule-based systems whose performance
degrades precipitously when confronted with tasks that go beyond what
has been explicitly encoded in the rule base, the analogical inference process
used in case-based systems generally degrades more gracefully, in much
the same fashion that human reasoning begins to fail but does not collapse
in the face of novelty (Waltz 1988).
After spelling out the motivations and characteristics of case-based
reasoning, we will illustrate the notion by applying it not to the task of
reconstructing the patterns in Luttwak's data, as in the previous section,
but to the task of reconstructing Luttwak's analysis and interpretation of
the data. Luttwak himself serves as an example of political intelligence at
work as he constructs the concept of coup d'etat from historical cases.

Instead of reproducing patterns in a body of data, we reproduce the process
by which political agents identify patterns or paradigms and reapply them
in new contexts. Ironically, the fundamental insight in CBR systems conforms
to the essentials of the information-processing model that Newell and Simon
introduced in the first decade of artificial intelligence (Newell and Simon
1972). What CBR offers in addition is a more explicit account of the objects
or structures in memory, their organization, their use in problem solving
and, as we will investigate in a later section, their acquisition. In short,
case-based reasoning can be approached as a refinement of the information-
processing model, 12 which holds that human beings solve problems by
assembling information in short-term memory that has been retrieved from
the virtually infinite store of long-term memory. Solving, for example, a
chess problem involves recognizing which of these structures in memory
are relevant to the current state of the game. Once recognized and retrieved,
the principles of analysis including problem decomposition and heuristic
search can be applied. The components of the information-processing model
are pictured in Figure 3.2.
The overlap with the basic conception of case-based reasoning is apparent
when this figure is juxtaposed to the flow of control typical of a CBR system,
as diagrammed in Figure 3.3, taken from Kolodner and Riesbeck (1989).
The two models differ in the perspective adopted: Newell and Simon focus
on the architecture of human problem solving, depicting the workspace of
short-time memory and its relation via a channel to a memory store. Kolodner
and Riesbeck, in their figure, focus on the functions of retrieving, adapting,
and evaluating cases in memory. What CBR adds is a more focused concern
with certain theoretical and computational issues such as that of how
structures in memory are indexed and how partial matches are handled
(Hammond 1989b).
Case-Based Reasoning: How It Works,

What Motivates Its Questions
"The case-based approach to planning is to treat planning tasks as
memory problems. Instead of building up new plans from scratch, a case-
based planner recalls and modifies past plans" (Hammond 1986: 287). The
case-based approach assumes that a basic form of learning takes place
from problem to problem or episode to episode. Rather than drawing
inferences by chaining together rules in deductive fashion, case-based systems
operate by matching the current problem against large structures in memory,
which comprise the system's store of previously solved problems or previously
encountered situations. What these structures are and how the memory is
organized to facilitate retrieval depends essentially on the characteristics of
72 Dwain Mefford
Figure 3.2: The Neweii-Simon Information Processing Model (Newel! and Simon, 1972;
Simon and Newel!, 1971) (Adapted from Davis and Olson, 1985, p. 242)
Internal
Long-term memory
I
Processor
Short-term memory
> Elementary processor >

Inpu t ou tput
Interpreter
t
External memory
the domain (Hammond 1986, 1987; Ashley and Rissland 1987, 1988; Mefford
1988d, 1989; Rissland and Ashley 1986).
What makes up the basic structures of memory? 13 The variation in ideas
on this subject is illustrated in the shifts in Roger Schank's thinking on
the question, from the master concept of "script" (Abelson 1973, 1978;
Schank and Abelson 1977; Schank 1982) to the idea of congeries of more
elemental scenelike units (Schank 1982: eh. 2). The question of what counts
as a "case" is equally problematic for Schank's students, and the students
of Schank's students, who are heavily represented among the proponents
of case-based reasoning (Carbonell 1981, 1986; Farrell 1987, 1988a, 1988b;
Hammond 1986, 1987; Kolodner 1983, 1984; Kolodner and Simpson 1984;
Kolodner, Simpson, and Sycara 1985; Sycara 1985a, 1985b, 1988a, 1988b).
It should come as no surprise that designers of systems intended to construct
legal briefs would seize upon precedent-based reasoning (Carter 1987; Woodard
1987).
Because the notion of "case" is so plastic, it is evident that what
distinguishes case-based reasoning from other forms of inference is not its
Figure 3.3: Case-Based Reasoning. Basic Flow of Control (From Kolodner and Riesbeck
1989, p. 2)
Problem
>~ Retrieve :
Best
>-
Statement Match
I
Cases
I Store : ~~
case and
Case
Memory
Prior
>I Adapt I
Indices Prior Adaptations Proposed
Outcomes Solution
New I
I Evaluate I<
case
Solution
Performance
data structure but something characteristic of the underlying reasoning

process. CBR is reasoning by analogy or example. Recognizing similarities
between one situation and another, or between a problem that has been
solved previously and a new problem for which a solution strategy is sought,
has been one of the principal topics in artificial intelligence from the
beginning. The task is to judiciously identify similarities and differences and
carry over from the one case what can be applied, plausibly, in the other.
Those insights, amplified by the first generation of problem-solving programs,
such as Evans's work with geometry analogy problems (1968), have propelled
a large branch of machine learning (Carbonnell 1982, 1986; Gentner 1980;
Winston 1975, 1979, 1980, 1981).
It is natural that this strategy or form of inference be put to use in other
domains, particularly those that are not closed and are, as a consequence,
not suitable for problem solving that uses resolution or similar deductive
techniques (Nilsson 1980; Genesereth and Nilsson 1987). Extending first-
order logic to encompass commonsense reasoning remains an unfulfilled
master goal of a significant branch of artificial intelligence (McCarthy and
Hayes 1969; McCarthy 1988; Genesereth and Nilsson 1987). In the absence
of a form of deduction that is adequate to open, ambiguous, and contradictory
domains, reasoning by analogy or reasoning by example provides a powerful
and flexible alternative inference mechanism, as has been recognized re-
peatedly in the history of logic (Peirce 1950). 14
74 Dwain Mefford
Model 3: Luttwak as a Case-Based Reasoning System

Luttwak's account of planning and executing a coup is a composite
portrait, assembled from roughly a dozen historical events, most prominently
Nasser's rise to power in 1952, the failed attempt by the French generals
in Algeria to depose de Gaulle in 1961, and the Portuguese revolt in 1975.
As apparent from Figure 3.1, the task of planning and executing a coup
consists of dozens, if not hundreds, of subtasks or problems. Each of these
subtasks can be analyzed using concrete historical examples. This is in fact
how Luttwak proceeds, e.g., developing the role of the "wait-and-see" factions
by analyzing the case of the attempted coup by the French generals (Luttwak
1979: 105-107). Developing a concept or structuring a problem on the
basis of its similarity to a concrete instance is a prime example of case-
based reasoning. Reasoning by example is perhaps the oldest and most
ubiquitous form of analysis and argument in the study of politics (Mefford
1988h).
Luttwak's discussion of the subtask of "neutralizing the police" can serve
as a limited example of case-based reasoning. Luttwak identifies three types
of police structure: that realized in the Paris Prefecture, that of the British
police force, and that of the FBI and other agencies in the United States.
Because the police and internal security agencies represent the first line of
defense for the state, defeating or eo-opting the police figures would be
among the first and foremost objectives for the organizers of a coup. To
accomplish this requires examining the functional organization of the police
in order to discover weak links that can be exploited. Because a significant
number of African and Middle Eastern regimes either inherited or adopted
organizational structures from the French or the British, Luttwak uses the
French and British police structures as surrogates for a general concept of
national police. Within Luttwak's overall theory of coups d'etat, the subtask
of "neutralizing the police" amounts to the problem of countermanding a
force with the characteristics of the French system.
Though the French police system is particularly extensive, its basic features
are shared by police forces in most of Africa, Asia and the Middle East. The
paramilitary element is usually present in the form of a "field force" attached
to the regular police, or else in the form of armored car units. The riot-control
element is reproduced in the special squads of Middle Eastern police forces
which, like Beirut's Squad 16, can be very effective in spite of their small
size (Luttwak 1979: 94).
For our purposes what is important is not the content of this analysis but
its logic. Concrete cases or examples are retrieved and used in lieu of
abstract concepts to address the problem at hand. 15 As Luttwak's analysis
of the Paris Prefecture illustrates, this reasoning by analogy can be quite
involved. In principle, a number of cases might be recognized as relevant,

which raises the issue of how to adjudicate conflicting claims as to which
example should prevail in informing the interpretation of the new situation
or problem. The process may also give rise to concepts that integrate
several cases from memory into a composite with no actual historical
precedent (Mefford 1988b).
Explanation-Based Learning
If a case-based reasoning system has the capacity to save the cognitive
work that it performs, then it becomes a learning system of at least a
rudimentary kind. If it saves that work in the form of a generalized plan,
then that system functions in the fashion of an explanation-based learning
system. This type of learning system differs fundamentally from that of
the learning programs that were first developed in artificial intelligence.
Those earlier learning systems, like behavioral learning models in general
(Sternberg 1963; Rapoport 1983: eh. 13), defined and modeled learning in
terms of adjustments to parameters. The parameters themselves were often
interpreted as statistical measures. The process of adjustment, therefore,
corresponds to Bayesian or some other form of estimation based on sampling
theory. Within artificial intelligence, and within computer science more
generally, this type of learning program belongs to the field and tradition
of statistical pattern recognition (Newell1983c: 197-199). Pattern recognition,
rooted in signal detection and information theory of the 1940s, gave rise
to what has been, retrospectively, called "similarity-based learning" or
"empirical learning" (Langley 1989; Schrodt, this volume).
In the early 1980s a second paradigm emerged. Sometimes called "analytic"
learning, the foremost example is "explanation-based learning." Learning in
this paradigm does not require large samples and is not expressed as
changes in the value of numerical parameters. What is learned is generally
a structure or configuration of information, e.g., a plan, and this can be
acquired on the basis of a single example. The contrast between similarity-
based and explanation-based learning can be overdrawn (Langley 1989), and
hybrid systems have been proposed. Nevertheless, the distinction is an
important one and the advent of explanation-based learning marks a significant
departure within formal learning theory.
Polya 's Maxim: Use What You Know

to Solve What You Do Not Know
As was the case for the essential intuitions in case-based reasoning, the
core ideas of explanation-based learning can be introduced through Polya's
concepts and examples, which are, in turn, rooted in the history of
76 Dwain Mefford
mathematics. Polya condenses the results of his study of heuristics into a

handful of maxims, including the following advice as to how to devise a
strategy for solving a problem: "Ask the question: Have you seen it before?
Or have you seen the same problem in a slightly different form?" (Polya
1945: xvi).
A classic application of this strategy in the earliest AI systems was
Simon, Newell, and Shaw's logic theorist (Newell and Simon 1956; Newell,
Shaw, and Simon 1958), which we encountered in an earlier section. Starting
with a list of rules for rewriting logical expressions, that program attempts
to prove theorems by construction. Once a proof has been constructed it
can be saved and reused in proofs of yet more theorems. Previously proved
theorems become cases in memory that can be modified or built upon to
solve new problems. This rudimentary form of learning via saving copies
of previous cognitive work is intrinsic to explanation-based learning, though
EBL as developed by Mitchell, Keller, and Kedar-Cabelli 1986, DeJong and
Mooney 1986, and Mooney and Bennett 1986 utilizes several additional
procedures.
Learning as the Retention of Plans

The strategy of saving and reusing proofs stems from the recognition
that it would be wasteful to discard them if there is any likelihood that the
same problem, or one like it, might be encountered again. In the STRIPS
project, once assembled into an executable plan the primitive operators
required to, say, move an object through a doorway, are bundled together
into a single operator. This new operator, called a "macro-operator," consists
of the conditions that must be true for the operators it contains to work,
plus the sequence of operators themselves. If the robot planner subsequently
encounters a goal that matches the consequent of the macro-operator, and
if the requirements of the macro-operator are met or can be met through
another round of problem solving, then the planner applies the macro-
operator and the goal is achieved.
The essentials of macro-operator acquisition are renewed in explanation-
based learning. The conception of what is acquired in the learning process
is broadened to include concepts in general rather than merely executable
plans. The connection between plans and the definition of a concept resides
in the notion that concepts, or an important subset of concepts, are essentially
classes or categories of objects. To "know" the concept is to acquire a
rule or formula for recognizing instances of the concept. This formula or
rule is a generalization of the proof that a given object is an instance of
the class or category.
Learning Concepts Essential to Interpersonal Relations

Explanation-based learning is partly rooted in the work with theorem
proving and macro-operator construction in robotic planning (DeJong 1983),
but it is also derivative of a practical and theoretical problem inherent in
the work on script-based systems that process natural language-how scripts
or schema are to be acquired. Schank and his students at Yale demonstrated
that computer programs can be written that respond appropriately to questions
about stylized situations, such as eating in restaurants, provided that they
have been programmed with these elaborate scriptlike structures. The
question immediately arises how such scripts are acquired. This is a theoretical
issue of the first rank. lt is also a practical issue because a Schankian
system that is tuned for one or a small number of hand-coded scripts, can,
at best, converse about menus and waitresses. The number and variety of
scripts that would be needed to handle free text is staggering. Because the
hand-coding of scripts is an extraordinarily labor-intensive activity that can
be measured at the rate at which Yale Ph.D. theses are produced, if Schank's
approach is to scale up, then a system based on Schank's principles must
be capable of extracting or devising new scripts on its own as it encounters
new situations. (DeJong's work on "explanatory schema acquisition," which
is one of the immediate forerunners of explanation-based learning, is an
effort to solve the problem of acquiring scripts and the related structures
needed by a script-based system by extracting structures of this sort from
situations [DeJong 1979, 1980].)
Explanation-Based Learning as a
Second Paradigm in Machine Learning
The conceptual and technical task of acquiring or learning schema differs
substantially from the learning task characteristic of much of the machine
learning prior to 1980. The conventional learning task, now sometimes called
similarity-based learning, involves constructing a discrimination function. As
the program is presented with examples of an unknown concept, it attempts
to construct that concept by assembling a set of predicates from the
underlying feature space (Michalski 1983). The task is essentially the same
as one that has been well studied by psychologists: The subject is presented
with a series of stimuli, say cards with colored shapes. The object is for
the subject to induce a concept, that is, to identify a set of descriptive
features that apply to the examples. Variations on this learning process
make use of negative examples or "near misses" (Winston 1975). In contrast
to this learning process, the learning of a schema may be based on a single
example or episode, e.g., in DeJong's work a single case of blackmail suffices
78 Dwain Me/ford
to establish that concept. Rather than induce the concept from a large
sample of instances, as in inductive learning, schema acquisition extracts
and generalizes the structure of interacting goals of the agents involved.
This learning process relies heavily on the knowledge base that the system
already possesses. In the case of DeJong's work (1980, 1981, 1982, 1983),
this knowledge is organized around the multilevel theory of plans and goals
that evolved from Abelson's earlier work with plans and scripts (1973, 1975),
as extended by Schank (Schank and Abelson 1977). 16
As this approach to learning was extended to domains other than
interpersonal relations, the notion of schema was replaced by the term
"explanation." lt is important to note that the learning of concepts in the
form of schema or explanations essentially amounts to reorganizing relations
or implications that are already implicit in a knowledge base. The new
concept is not "induced" in the sense that it is added to the knowledge
base from some external source. The utility in this form of learning lies in
the efficiency with which the knowledge it contains can be subsequently
applied. What the system in effect does is to save a generalized copy of
the cognitive work, e.g., the steps that it went through in identifying an
object as an instance of a concept. This is precisely what occurs when a
theorem-proving system records and reuses a theorem or lemma that it
has proved. In a literal sense there is no new information, but the information
is now explicit where it had been previously only implicit in the system's
knowledge base. Learning in this paradigm, therefore, is best conceived as
the reorganization of the knowledge base on the occasion of, and for the
purpose of, more effectively executing some task.
EBL as of the mid-1980s can be characterized as a process consisting
of three or four steps. First, unlike much of conventional learning theory
in which the program lacks knowledge or any sort, in EBL it is assumed
that the program is equipped with a collection of facts and relationships
that constitute a "theory" of some domain, e.g., the behavior of liquids or
physical objects. Second, the concept to be acquired is described in terms
of the function or goal that it serves. The concept of a "cup" is initially
described in terms of its function as a vessel for holding liquids. How this
function is realized in actual examples is for the program to determine as
it fteshes out the idea. The third step or component in EBL is, therefore,
the training example or instance. The task for the system is to construct
an "explanation" of how the properties of this example satisfy or accomplish
the functional definition of the concept, e.g., how a particular cup realizes
the function of holding or conveying liquids. In the EBL programs that have
been written to date, the learned concept or explanation takes the form of
a proof. That is, the concept that is constructed is the sequence of steps
that determine whether an object is or is not an instance of the concept.
Figure 3.4: Concept of a "Cup" in the Form of a Proof That an Object Meets the
Functional Requirements of a Cup (Modified from DeJong, 1988, p. 56)
cup (_i.BJX)
drinkable-from (OBJX)
~iftabll(OBJX) Liquid-container(OBJX), Stable,(OBJX)
I I I
Has-part(OBJX,OBJZ)
Weight(OBJ!,light) Isa(OBJZ,
FLAT-BOTTOM)
Isa(OBJY,HANDLE)
Has-part(OBJX,
CONC12)
Has-part(OBJX,OBJY) Orientation(CONC12,
UPWARD)
Isa(CONC12,CONCAVITY)
As noted, the steps in explanation-based learning are often presented

using the example of a concept of a "cup," which has been a favorite in
machine learning since Winston used it (Winston, Binford, Katz, and Lowry
1983). The concept acquired takes the form of proof structure like that in
Figure 3.4. Given that the program knows something about objects, such
as handles and open containers, its task is to construct a concept of "cup"
by assembling this information. lt proceeds by inspecting an example of a
cup, consisting of a configuration of relationships and properties, which it
augments and generalizes to form the concept. This is accomplished in the
following fashion. Let us assume that an object can serve functionally as
a cup if it can hold liquids and if people can drink from it. Let us further
assume that we already know that any upward-pointing open vessel can
hold a liquid, and that a person can drink from an object if it has a usable
handle, and a usable handle is one that does not get overly hot when
attached to a vessel, etc., etc. Faced with an object that is putatively a
cup, the learning system will attempt to prove, on the basis of what it
already knows, that the object is functionally equivalent to a cup. If it
succeeds in constructing this proof, then the proof, properly generalized,
becomes the concept that is added to the knowledge base. This concept
can be read "procedurally": to show that some object is a "cup," show
80 Dwain Me/ford
Figure 3.5: Explanation of the Concept of Suicide in the Form of a Proof Tree (Modified
from DeJong and Mooney, 1986, p. 159)
I
HATE(x,x) POSSESS(x,objl)
I
WEAPON(objl)
I
DEPRESSED(X)
I
BUY(x,objl)
I
GUN(objl)
that it is stable, liftable, and an open vessel. To show that an object is

stable, show that it has a flat bottom, etc.
Switching examples for the purpose of tying EBL to schema acquisition,
a scriptlike pattern of action can be represented in the same form as the
cup example. The concept or schema of "suicide" is diagrammed in Figure
3.5. Abstracted from a concrete example of suicide, the generalized concept
says that anyone who is depressed and who buys a gun may intend to use
that gun to kill himself.
EBL, like other learning processes investigated in artificial intelligence
(Laird, Newell, and Rosenbloom 1987; Laird, Rosenbloom, and Newell 1986;
Rosenbloom and Laird 1986), is still very much in an experimental stage.
It has been applied to at most a small number of examples. Nevertheless,
the object is to devise a learning procedure that will progressively improve
the performance of the system by acquiring new concepts. Lenat's AM
(Lenat 1977, 1982), though not an EBL system, is an example of a heuristic
learning program set loose in a virtually infinite domain.
Model 4: Political Learning and Path-Dependent Processes

Elsewhere we have constructed EBL models of change in U.S. foreign
policy (Mefford 1988b, 1988e, 1988g, 1989, forthcoming). Our claim is that
explanation-based learning, involving the acquisition of concepts from small
samples of events, or even from single examples, is sufficient to generate
the type of qualitative, path-dependent change witnessed in the evolution
of foreign policy over time. Episodes, elevated to the status of concepts,
are in effect recycled in subsequent planning and counterplanning. Concept
acquisition of this order can alter the structure of the competitive "games"
that constitute the dynamics of international politics (Mefford 1988f, 1988g).
The question is not whether these historically derived concepts are valid-
valid or not they figure in the political imagination, affecting how courses
of action are conceived and evaluated. There is a complex interaction between
the concepts used to interpret situations and the impact those situations
may exert, after the fact, on the stock of effective political concepts. The
suggestion is that the language in which policy is formulated itself carries
the imprint of previous experience and is continually subject to revision.
This language-based process of political change is path-dependent in two
respects. Not only do singular events become part of the working set of
concepts that shape policy-and in this way later policy is dependent on
what occurred earlier in the historical path-but what is extracted from an
event and how that lesson is integrated into the working corpus of concepts
is a function of the state of political thinking at the time of the event. The
order of experience matters because each episode has the potential to alter
not only the "stock" of concepts but also the way in which the new example
is incorporated into the body of resident knowledgeY The universe of
argument or debate that that language will support at a point in time is
qualitatively different by virtue of the insertion or modification of new
concepts. Vietnam is perhaps the most recent and vivid example. After the
historical experience of Vietnam, "Vietnam" exists as a category that enters
into how Americans think about military intervention. The public debate
over U.S. policy in Central America is formulated in the shadow of that
image and all that it conveys to a generation.
To move from the conception of a process to a full theory of this type
of political change requires considerable formal and empirical development.
On the formal side, it is apparent that the process of extracting concepts
from historical cases involves a much more elaborate selection and filtering
process than the ones that have been programmed in existing EBL systems.
On the empirical side, major effort is needed to "calibrate" the learning
process. We must closely inspect what distinguishes those episodes that
alter the very language of foreign policy from the mass of routine events
that have little or no effect on the working categories used to formulate
policy.
To describe this process of conceptual change as one of "learning"
perhaps invites misunderstanding. We are not accustomed to think of learning
as a process that transpires at the level of the language of a society. Yet,
formally, the process of acquiring and revising political concepts in the face
of events very much resembles the process of concept acquisition embodied
in the EBL algorithms. That is, the effective political language, by which
we mean the concepts and distinctions used to identify policy options, often
does change through an abrupt, single-example type of learning like that
epitomized in EBL programs. For example, there is a qualitative difference
between the reasoning that characterized strategic relations before October
1962 and the reasoning that has prevailed since (Mefford 1989b). How this
can happen is a fundamental scientific question for students of international
relations. We are suggesting that the underlying process can be conceptualized
82 Dwain Mefford
as formally identical to the conception of system change embodied in

explanation-based learning. 18
Notes
1. The conceptual fit between the two is apparent in a number of the papers in
the proceedings of the most recent case-based reasoning workshop, including the
opening statement titled "Case-Based Reasoning from DARPA: Machine Learning
Program Plan" (Hammond 1989b).
2. On the rhetorical power of the phrase "explained variance," and on the
confusions it engenders, see Donald McCiosky, "The Rhetoric of Significance Tests"
(McCiosky 1985).
3. Though it stops with applications as of late 1986, the best overview of
applications of artificial intelligence by political scientists is Mallery 1988. Other
surveys include the chapters in Cimbala 1987, especially Andriol 1987.
4. On the question whether a distinction should be made, Davis flatly states that
rule-based and production-based systems are identical in their essentials (Davis and
King 1976). But Charniak, Riesbeck, and McDermott (1980: 222-223) distinguish
between the two on the grounds that productions are unlimited in the actions that
they can execute, e.g., they can perform input-output functions, whereas rules are,
or should be, constrained to the making of inferences under the control of a deductive
procedure.
5. Expanding from his initial interest in psycholinguistics, since 1978 John
Anderson has constructed a general theory of the architecture of cognition based
on the concept of production systems. Since about 1978, that theory has been
realized in a collection of computer programs, ACT* and its successors. G. R.
Boynton has applied and adapted elements of Anderson's theory in the effort to
model the process of decision making and the fiow of information in congressional
and Senate committees (Boynton 1987, 1989, this volume).
6. For an inventory of the most important of these systems presented in the
form of a family tree, see Figure 1-3 in Buchanan and Shortliffe 1984: 11. Hayes-
Roth, Waterman, and Lenat present a similar genealogy of the principle rule-based
systems in Figure 1.1, p. 8 of Hayes-Roth, Waterman, and Lenat 1983.
7. The best accounts of how to engineer rule-based systems is Buchanan and
Shortliffe 1984; Hayes-Roth, Waterman, and Lenat 1978, 1983. For the current state
of rule-based systems see Hayes-Roth 1984 and articles in the journal IEEE Expert.
8. Implementation, for example in OPS5 (Brownston, Farrell, Kant, and Martin
1985) or Prolog (Hammond 1982, 1984) is immaterial. We are concerned with
questions at a conceptual level higher than the choice of programming language or
shell.
9. The six generic tasks are characterized by their input-output behavior, their
representations, and their control regimes. The first three, hierarchical classification,
hypothesis matching, and knowledge-directed information passing, designate functions
characteristic of diagnostic/classification expert systems. The second three, object
synthesis by plan selection and refinement, state abstraction, and abductive assembly
of hypotheses, have been added since the first group were abstracted from Chan-
drasekaran's program MDX. These additions parallel a number of Clancy's efforts

(Chandrasekaran 1986: 24).
10. Within organizational theory, March and Cohen have pursued issues of
problemsolving and planning at the institutional level (Cohen, March, and Olsen
1972; Cohen 1981, 1984).
11. For an argument that AI modeling in political science should set itself the
weaker goal of describing regularities and making predictions, but not attempting
to reproduce real processes, see Schrodt 1988b.
12. For an account of case-based reasoning that presents it as a more radical
departure within artificial intelligence, see Riesbeck and Schank 1989.
13. For a survey of the impact of memory research, particularly Tulving's
distinction between "episodic" and "semantic" memory on artificial intelligence and
cognitive science, see Mefford 1987c.
14. For a fuller examination of what is involved in reasoning by example, plus
the history of this notion from Aristotle's Prior Analytics forward, see Mefford 1988h.
15. Bloomfleld's CASCON has many of the properties of a case-based reasoning
system, though it does not in fact reason about the cases. It retrieves cases that
match the properties of a dispute. The system leaves the task of drawing inferences
to the user, and, as far as I am aware, has no learning component. The result of
a session with CASCON is not added to the data base of cases. CASCON is not,
in itself, a cognitive model, though it is built upon keen intuitions as to how foreign
policy professionals use information to structure decisions (Bioomfleld and Beattie
1971; Bloomfleld 1986).
16. For an account of scripts, plans, and goal structures see Mefford 1987c.
Within political science the first application of Abelson-Schank concepts is to be
found in Alker 1975.
17. The algorithm used in Mefford 1988e is taken from Mooney 1986. There is
an error in the scope of the generalization of the variables in Mitchell's algorithm
(Mitchell, Keller, and Kedar-Cabelli 1986), see DeJong and Mooney 1986. For EBL
as a metaprogram in Prolog, see Kedar-Cabelli and McCarty 1987 and Kodratoff
1988.
18. There are, of course, other processes involved, some of which are countervailing.
Explanation-based learning is rapidly evolving in the face of new challenges, which
include such issues as the question of the costs of indiscriminate learning. This
involves, among other questions, that of the trade-off between the expense of
maintaining and searching a large memory versus constructing concepts "as needed"
(Tambe and Newell 1988).
Bibliography
Abelson, Robert P. 1973. "The Structure of Belief Systems." In Computer Models
of Thought and Language, ed. Roger C. Schank and Kenneth Mark Colby. San
Francisco, CA: W.H. Freeman, pp. 287-339.
- - - . 1975. "Concepts for Representing Mundane Reality in Plans." In Repre-
sentation and Understanding Studies in Cognitive Science, ed. Daniel G. Bobrow
and Allan Collins. New York: Academic Press, pp. 273-309.
84 Dwain Me/ford
- - - . 1978. "Scripts." Invited address to the Midwestern Psychological Association.

New Haven, CT: Yale University Press.
- - - . 1981. "Constraint, Construal, and Cognitive Science." Paper delivered at
the Third Annual Cognitive Science Conference, Berkeley, CA.
Aguilar, L., ed. 1981. Operation ZAPATA: The "U/trasensitive" Report and Testimony
of the Board of Inquiry on the Bay of Pigs. Frederick, MD: Aletheia Books.
Alker, Hayward R., Jr. 1975. "Polimetrics: Its Descriptive Foundations." In Handbook
of Political Science, ed. Fred I. Greenstein and Nelson W. Polsby. Reading, MA:
Addison-Wesley, vol. 7, pp. 139-210.
- - - . 1981. "From Political Cybernetics to Global Modeling." In From National
Development to Global Community, ed. Richard L. Merritt and Bruce M. Russett.
London: George Alien & Unwin, pp. 353-378.
- - - . 1984. "Historical Argumentation and Statistical Inference: Towards More
Appropriate Logics for Historical Research." Historical Methods 17, 3: 164-173.
Alker, Hayward R., Jr., James P. Bennett, and Dwain Mefford. 1980. "Generalized
Precedent Logics for Resolving Insecurity Dilemmas." International Interactions
7, 2: 165-206.
Alker, Hayward R., Jr., and William J. Greenberg. 1977. "On Simulating Collective
Security Regime Alternatives." In Thought and Action in Foreign Policy, eds.
Mathew Bonham and Michael Shapiro. Basel: Birkhauser.
Anderson, John R. 1976. Language, Memory, and Thought. New York: Wiley, 1976.
___ . 1983. The Architecture of Cognition. Cambridge, MA: Harvard University
Press.
Andriole, Stephen J. 1987. "Artificial Intelligence Based Simulations of Foreign Policy
Decision Making." In Artificial Intelligence and National Security, ed. Steven J.
Cimbala. Lexington, MA: Lexington Books, pp. 21-34.
Ashley, Kevin D., and Edwina L. Rissland. 1987. "Compare and Contrast, A Test
of Expertise." Proceedings of the Sixth National Conference on Artificial Intelligence.
Menlo Park, CA: American Association for Artificial Intelligence.
___ . 1988. "A Case-Based Approach to Modeling Legal Expertise." IEEE Expert
3, 3: 70-77.
Bennett, James P., and Hayward R. Alker, Jr. 1977. "When National Security Policies
Bred Collective Insecurities." In Problems of World Modeling, ed. Karl W. Deutsch,
Bruno Fritsch, Helio Jaguaribe, and Adrei S. Markovits. Cambridge, MA: Ballinger,
pp. 219-302.
Bloomfield, Lincoln P. 1986. "Analytical Rigor and Real-World Conflicts: The Death
and Life of CASCON." Paper prepared for the annual meeting of the International
Studies Association, Anaheim, CA.
Bloomfield, Lincoln P., and R. Beattie. 1971. "Computers and Policy-Making: The
CASCON Experiment." Journal of Conflict Resolution 15: 33-46.
Boyd, Richard N. 1984. "The Current Status of Scientific Realism." In Scientific
Realism, ed. Jarrett Leplin. Berkeley, CA: University of California Press, pp. 41-
82.
Boynton, G. Robert. 1987. "Committee Law Making." Paper presented at the annual
meeting of the American Political Science Association, Chicago.
___ . Forthcoming. "Communication and Cognition: News Accounts of International
Affairs." International Studies Quarterly.
Brown, John Seely, Alien Collins, and G. Harris. 1978. "Artificial Intelligence and
Learning Strategies." In Learning Strategies, ed. Harry O'Neill. New York: Academic
Press.
Brownston, Lee, Robert Farrell, Elaine Kant, and Nancy Martin. 1985. Programming
Expert Systems in OPS5: An Introduction to Rule-Based Programming. Reading,
MA: Addison-Wesley.
Buchanan, Bruce G., and Edward H. Shortliffe. 1984. Rule-Based Expert Systems:
The MYCIN Experiments of the Heuristic Programming Project. Reading, MA:
Addison-Wesley.
Buchanan, Bruce G., G.L. Sutherland, and Edward A. Feigenbaum. 1969. "Heuristic
DENDRAL: A Program for Generating Explanatory Hypotheses in Organic Chem-
istry." In Machine Intelligence 4, eds. Bernard Meltzer and Donald Michie. New
York: American Elsevier.
Bull, Hedley. 1984. "Intervention in the Third World." In Intervention in World Politics,
ed. Hedley Bull. Oxford: Clarendon Press, pp. 135-156.
Bylander, Tom C., Sanjoy Mittal, and B. Chandrasekaran. 1983. "CSRL: A Language
for Expert Systems for Diagnosis." Proceedings of the Eighth International Joint
Conference on Artificial Intelligence. Los Altos, CA: Morgan Kaufmann, pp. 218-
221.
Bylander, Tom C., and Sanjoy Mittal. 1986. "CSRL: A Language for Classificatory
Problem Solving and Uncertainty Handling." A/ Magazine 7, 2: 66-77.
Carbonell, Jaime G., Jr. 1981. "Counterplanning: A Strategy-Based Model of Adversary
Planning in Real-World Situations." Artificial Intelligence 16: 295-329.
___ . 1982. "Experiential Learning in Analogical Problem Solving." Proceedings
of the Second National Conference on Artificial Intelligence. Menlo Park, CA:
American Association for Artificial Intelligence, pp. 168-171.
___ . 1986. "Derivational Analogy: A Theory of Reconstructive Problem Solving
and Expertise Acquisition." In Machine Learning: An Artificial Intelligence Ap-
proach, eds. R.S. Michalski, J.G. Carbonell, and T.M. Mitchell. Los Altos, CA:
Morgan Kaufmann.
Carter, Lief H. 1987. "Legal Reasoning." In Encyclopedia of the American Judicial
System: Studies of the Principle Institutions and Processes of Law, ed. Robert
J. Janosik. New York: Charles Scribner's Sons, vol. 2: 875-889.
Chandrasekaran, B. 1983. "Towards a Taxonomy of Problem-Solving Types." AI
Magazine 4, 1(Winter/Spring): 9-17.
- - - . 1984. "Expert Systems: Matching Techniques to Tasks." In Artificial
Intelligence Applications for Business, ed. W. Reitman. Norwood, NJ: Ablex, pp.
116-132.
- - . 1986. "Generic Tasks in Knowledge-Based Reasoning: High-Level Building
Blocks for Expert System Design." IEEE Expert 1, 3: 23-30.
- - - . 1987. "Towards a Functional Architecture for Intelligence Based on Generic
Information Processing Tasks." Proceedings of the Tenth International Joint
Conference on Artificial Intelligence. Palo Alto, CA: Morgan Kaufmann, pp. 1183-
1192.
Chandrasekaran, B., and S. Mittal. 1982. "Deep versus Compiled Knowledge Ap-
proaches to Diagnostic Problem-Solving." Proceedings of the National Conference
on Artificial Intelligence. Los Altos, CA: Morgan Kaufmann. pp. 349-354.
86 Dwain Me/ford
Charniak, Eugene, Christopher K. Riesbeck, and Drew V. McDermott. 1980. Artificial

Intelligence Programming. Hillsdale, NJ: Lawrence Erlbaum.
Cimbala, Stephen, ed. 1987. Artificial Intelligence and National Security. Lexington,
MA: Lexington Books.
Clancey, William J. 1983. "The Epistemology of a Rule-Based Expert System-A
Framework for Explanation." Artificial Intelligence 20: 215-251.
___ . 1985. "Representing Control Knowledge as Abstract Tasks and Metarules."
Stanford University, Knowledge Systems Laboratory Working Papers No. KSL 85-
16, April 1985.
___ . 1986. "From Guidon to Neomycin and Heracles in Twenty Short Lessons."
AI Magazine 7, 3(August): 40-60.
_ _ . 1987. Knowledge-Based Tutoring. The GUIDON System. Cambridge, MA:
MIT Press, 1987.
Cohen, Michael D. 1981. "The Power of Parallel Thinking." Journal of Economic
Behavior and Organization 2: 285-306.
___ . 1984. "Conflict and Complexity: Goal Diversity and Organizational Search
Effectiveness." American Political Science Review 78: 435-451.
Cohen, Michael D., James G. March, and Johan P. Olsen. 1972. "A Garbage Can
Model of Organizational Choice." Administrative Science Quarterly 17: 1-25.
Cook, Blanche Wiesen. 1981. The Declassified Eisenhower. A Divided Legacy. Garden
City, NY: Doubleday.
Crecine, J. Patrick. 1969. Governmental Problem Solving. Chicago: Rand McNally.
David, Paul. 1985. "Clio and the Economics of QWERTY." American Economic
Review Proceedings 75: 332-337.
Davis, Gordon B., and Margrethe H. Olson. 1984. Management Information Systems.
2d edition. New York: McGraw-Hill.
Davis, Randall. 1979. "Interactive Transfer of Expertise." Artificial Intelligence 12:
121-157.
___ . 1980. "Meta-rules: Reasoning about Control." Artificial Intelligence 15: 179-
222.
___ . 1982. "Expert Systems: Where Are We? And Where Do We Go From
Here?" AI Magazine 3, 2: 3-22.
___ . 1984a. "Amplifying Expertise with Expert Systems." In The AI Business:
The Commercial Uses of Artificial Intelligence, ed. Patrick H. Winston and Karen
A. Prendergast. Cambridge, MA: MIT Press, pp. 17-40.
___ . 1984b. "Diagnosis Based on Structure and Function: Paths of Interaction
and the Locality Principle." Artificial Intelligence 24: 347-410.
Davis, Randall, Bruce G. Buchanan, and Edward H. Shortliffe. 1977. "Production
Rules as a Representation for a Knowledge-Based Consultation Program." Artificial
Intelligence 8: 15-45.
Davis, Randall, and Jonathan J. King. 1976. "An Overview of Production Systems."
In Machine Intelligence, eds. E. W. Elcock and D. Michie. New York: John Wiley,
pp. 300-332.
___ . 1984. "The Origin of Rule-Based Systems in Al." In Rule-Based Expert
Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project,
eds. Bruce G. Buchanan and Edward H. Shortliffe. Reading, MA: Addison-Wesley,
pp. 20-52.
DeJong, Gerald. 1979. "Prediction and Substantiation: A New Approach to Language

Processing." Cognitive Science 3: 251-273.
___ . 1980. "Automatic Schema Acquisition in a Natural Language Environment."
Proceedings of the First Annual National Conference on Artificial Intelligence.
Menlo Park, CA: American Association for Artificial Intelligence, pp. 410-413.
___ . 1981. "Generalizations Based on Explanations." Proceedings of the Inter-
national Joint Conference on Artificial Intelligence. Los Altos, CA: Morgan Kauf-
mann, pp. 67-69.
___ . 1982. "Automatic Schema Acquisition in a Natural Language Environment."
1982. Proceedings of the National Conference on Artificial Intelligence. Menlo
Park, CA: American Association for Artificial Intelligence, pp. 410-413.
___ . 1983. "Acquiring Schemata Through Understanding and Generalizing Plans."
Proceedings of the Eighth International Joint Conference on Artificial Intelligence
Los Altos, CA: Morgan Kaufmann. vol. 1: 462-464.
___ . 1988. "An Introduction to Explanation-based Learning." In Exploring Artificial
Intelligence: Survey Talks from the National Conferences on Artificial Intelligence,
ed. Howard E. Shrobe. Palo Alto, CA: Morgan Kaufmann, pp. 45-81.
DeJong, Gerald, and Raymond Mooney. 1986. "Explanation-Based Learning: An
Alternative View." Machine Learning 1, 2: 145-176.
Duffy, Gavan. 1988. "Theory and Text: How Textual Analysis Can Help Build
International Relations Theory." Paper presented at the annual meeting of the
International Studies Association, St. Louis.
Duffy, Gavan, and John C. Mallery. 1986. "RELATUS: An Artificial Intelligence Tool
for Natural Language Modeling." Paper presented at the annual meeting of the
International Studies Association, Anaheim, CA.
Durfee, Edmund H. 1986. "Incremental Planning to Control a Blackboard-Based
Problem-Solver." Proceedings of the National Conference on Artificial Intelligence.
Palo Alto, CA: Morgan Kaufmann, pp. 58-64.
Durfee, Edmund H., and Victor R. Lesser. 1987. "Using Partial Global Plans to
Coordinate Distributed Problem Solvers." Proceedings of the International Joint
Conference on Artificial Intelligence. Palo Alto, CA: Morgan Kaufmann, pp. 875-
883.
Durfee, Edmund H., Victor R. Lesser, and Daniel D. Corkill. 1987. "Coherent Cooperation
Among Communicating Problem Solvers." IEEE Transactions on Computers. C-
36: 1275-1291.
Ernst, George W., and Alien Newell. 1969. GPS: A Case Study in Generality and
Problem Solving. New York: Academic Press.
Evans, Thomas G. 1968. "A Program for the Solution of Geometric-Analogy Intelligence
Test Questions." In Semantic Information Processing, ed. Marvin Minsky. Cam-
bridge, MA: MIT Press, pp. 271-353.
Farrell, Robert. 1987. "Intelligent Case Selection and Presentation," Proceedings of
the Eleventh International Joint Conference on Artificial Intelligence. Los Altos,
CA: Morgan Kaufmann.
- - - . 1988a. "Facilitating Self-Education by Questioning Assumptive Reasoning
using Paradigm Cases." In Proceedings, Case-Based Reasoning Workshop, ed.
Janet Kolodner. San Mateo, CA: Morgan Kaufmann.
88 Dwain Mefford
- - - . 1988b. "Facilitating Self-Education by Questioning Assumptive Reasoning,"

Proceedings of the Seventh National Conference on Artificial Intelligence. San
Mateo, CA: Morgan Kaufmann.
Feigenbaum, Edward A. 1963. "The Simulation of Natural Learning Behavior." In
Computers and Thought, eds. E.A. Feigenbaum and J. Feldman. New York:
McGraw-Hill.
- - - . 1977. "The Art of Artificial Intelligence: I. Themes and Case Studies of
Knowledge Engineering." Proceedings of the International Joint Conference on
Artificial Intelligence. Pittsburgh, PA: Carnegie-Mellon University, pp. 1014-1029.
Feigenbaum, Edward A., Bruce G. Buchanan, and Joshua Lederberg. 1971. "On
Generality and Problem Solving: A Case Study Using the DENDRAL Program."
In Machine Intelligence, eds. B. Meltzer and D. Michie. Edinburgh: Edinburgh
University Press, vol. 6, pp. 165-190.
Feller, William. 1950. An Introduction to Probability Theory and Its Applications,
vol. 1. New York: Wiley.
Fikes, Richard E., and Nits J. Nilsson. 1971. "STRIPS: A New Approach to the
Application of Theorem Proving to Problem Solving." Artificial Intelligence 2, 3/
4: 189-208.
Fikes, Richard E., P.E. Hart, and Nits J. Nilsson. 1972. "Learning and Executing
Generalized Robot Plans." Artificial Intelligence 3, 4: 251-288.
Floyd, R. 1961. "A Descriptive Language for Symbol Manipulation." Journal of the
Association for Computing Machinery 8: 579-584.
Genesereth, Michael R., and Nits J. Nilsson. 1987. Logical Foundations of Artificial
Intelligence. Los Altos, CA: Morgan Kaufmann.
Gentner, Dedre. 1980. "The Structure of Analogical Models in Science." Technical
Report No. 4451. Cambridge, Mass.: Bolt, Beranek & Newman, July 1980.
George, Alexander L., and Richard Smoke. 1989. "Deterrence and Foreign Policy."
World Politics 61, 2: 170-182.
Georgeff, Michael P. 1987. "Planning." In Annual Review of Computer Science 2,
eds. Joseph F. Traub, Barbara J. Grosz, Butler W. Lampson and Nits J. Nilsson.
Palo Alto, CA: Annual Reviews, Inc., 359-400.
Gilbert, F. 1965. Machiavelli and Guicciardini. Princeton, NJ: Princeton University
Press.
Goel, Ashok, and B. Chandrasekaran. 1987. "JESSE: A Conceptual Model of Political
Decision Making." Paper presented at the annual meeting of the Midwest Artificial
Intelligence and Cognitive Science Society, Chicago.
Goel, Ashok, B. Chandrasekaran and Donald Sylvan. 1987 "JESSE: An Information
Processing Model of Policy Decision Making." Proceedings of the Third Annual
Expert Systems in Government Conference, sponsored by IEEE Computer Society,
Washington, D.C., October 1987, pp. 178-187.
Haas, Ernst B. 1982. "Words Can Hurt You: Or, Who Said What to Whom about
Regimes." International Organization 36, 2: 207-243.
Halliday, M.A.K. 1975. Learning How to Mean: Explorations in the Development of
Language. London: Edward Arnold.
Hammond, Kristian J. 1986. "Learning to Anticipate and Avoid Planning Problems
Through the Explanation of Failures." Proceedings of the National Conference
on Artificial Intelligence. Los Altos, CA: Morgan Kaufmann, pp. 556-560.
___ . 1987. "Explaining and Repairing Plans that Fail." Proceedings of the Tenth
International Joint Conference on Artificial Intelligence. San Mateo, CA: Morgan
Kaufmann, vol. 1: 109-114.
___ . 1989a. Case-Based Planning: Viewing Planning as a Memory Task. New
York: Academic Press.
___ , ed. 1989b. Case-Based Reasoning: Proceedings of a Workshop on Case-
Based Reasoning. San Mateo, CA: Morgan Kaufmann.
Hammond, P. 1982. "Appendix to Prolog: A Language for Implementing Expert
System." Machine Intelligence, eds. Bernard Meltzer and Donald Michie, vol. 10.
New York: Elsevier, pp. 471-475.
___ . 1984. "Micro-PROLOG for Expert Systems." In Micro-PROLOG: Program-
ming in Logic, eds. K.L. Clark and F.G. McCabe. Englewood Cliffs, NJ: Prentice-
Hall International, pp. 294-319.
Harmon, Paul. 1989. "U.S. Expert Systems Building Tools." Expert Systems Strategies
5, 8: 1-8.
Harre, Rom. 1986. Varieties of Realism: A Rationale for the Natural Sciences. Oxford:
Basil Blackwell.
Hayes, Patrick J. 1977. "In Defense of Logic." Proceedings of the Fifth International
Joint Conference on Artificial Intelligence. Pittsburgh, PA: Carnegie-Mellon Uni-
versity, pp. 559-565.
Hayes-Roth, Frederick. 1985. "Rule-Based Systems." Communications of the ACM
28, 9(September): 921-932.
Hayes-Roth, Frederick, Donald A.Waterman, and Douglas B. Lenat. 1978. "Principles
of Pattern-Directed Inference Systems." In Pattern-Directed Inference Systems,
eds. Donald A. Waterman and Frederick Hayes-Roth. New York: Academic Press,
pp. 577-601.
___ . 1983. Building Expert Systems. Reading, MA: Addison-Wesley.
Hewitt, Carl. 1971. "Description and Theoretical Analysis (Using Schemas) of
PLANNER: A Language for Proving Theorems and Manipulating Models in a
Robot." Doctoral dissertation, Massachusetts Institute of Technology.
lmmerman, Richard H. 1982. The CIA in Guatemala: The Foreign Policy of Intervention.
Austin, TX: University of Texas Press.
Job, Brian L., and Doug Johnson. 1986. "Model of U.S. Foreign Policy Decision
Making: The U.S. and the Dominican Republic, 1961-1965." Paper presented at
the annual meeting of the International Studies Association, Anaheim, CA.
Job, Brian L., Douglas Johnson, and Eric Selbin. 1987. "A Multi-Agent, Script-Based
Model of U.S. Foreign Policy Towards Central America." Paper prepared for the
annual meeting of the American Political Science Association, Chicago.
Kedar-Cabelli, Smadar T., and L. Thorne McCarty. 1987. "Explanation-Based Gen-
eralization as Resolution Theorem Proving." In Proceedings of the Fourth Inter-
national Workshop on Machine Learning, ed. Pat Langley. Los Altos, CA: Morgan
Kaufmann, pp. 383-389.
Kodratoff, Yves. 1988. Introduction to Machine Learning. San Mateo, CA: Morgan
Kaufmann.
Kolodner, Janet L. 1980. "Organizing Memory and Keeping lt Organized." Proceedings
of the First Annual National Conference on Artificial Intelligence, Menlo Park,
CA: AAAI.
90 Dwain Me/ford
___ . 1983. "Maintaining Organization in a Dynamic Long Term Memory." Cognitive

Science 7: 243-280.
___ . 1984. Retrieval and Organizational Strategies in Conceptual Memory: A
Computer Model. Hillsdale, NJ: Lawrence Erlbaum.
___ . 1987. "Extending Problem Solver Capabilities Through Case-Based Inference."
Proceedings of the Fourth International Workshop on Machine Learning. University
of California, lrvine, June 22-25, 1987. Los Altos, CA: Morgan Kaufmann, pp.
167-178.
___ , ed. 1988. Proceedings of the DARPA Workshop on Case-Based Reasoning.
Palo Alto, CA: Morgan Kaufmann.
Kolodner, Janet L., and Chris Riesbeck. 1989. "Case-Based Reasoning." Tutorial:
MA2. Eleventh International Joint Conference on Artificial Intelligence. San Mateo,
CA: AAAI.
Kolodner, Janet L., and Robert L. Simpson. 1984 "Experience and Problem Solving:
A Framework." Proceedings of the Ninth Annual Conference of the Cognitive
Science Society. Hillsdale, NJ: Lawrence Erlbaum.
Kolodner, Janet L., R.L. Simpson, and K. Sycara. 1985. "A Process Model of Case-
Based Reasoning in Problem Solving." Proceedings of the Ninth International
Joint Conference on Artificial Intelligence. Los Altos, CA: Morgan Kaufmann, pp.
284-290.
Kowalski, Robert A. 1979. Logic for Problem Solving. Amsterdam: North-Holland.
Laird, John E., Alien Newell, and Paul S. Rosenbloom. 1987. "SOAR: An Architecture
for General Intelligence." Artificial Intelligence 33: 1-64.
Laird, John E., Paul S. Rosenbloom, and Alien Newell. 1986. Universal Subgoaling
and Chunking: The Automatic Generation and Learning of Goal Hierarchies.
Norwell, MA: Kluwer Academic.
Lakoff, George. 1987. Women, Fire, and Dangerous Things: What Categories Reveal
about the Mind. Chicago: University of Chicago Press.
Langley, Pat. 1989. "Unifying Themes in Empirical and Explanation-Based Learning."
Proceedings of the Sixth International Workshop on Machine Learning. San Mateo,
CA: Morgan Kaufmann. pp, 2-4.
Lebow, Richard Ned, and Janice Gross Stein. 1989. "Rational Deterrence Theory: I
Think, Therefore I Deter." World Politics 61, 2: 208-224.
Lehnert, Wendy G. 1978. The Process of Question Answering. Hillsdale, NJ: Lawrence
Erlbaum.
Lenat, Douglas B. 1977. "Automated Theory Formation in Mathematics." Proceedings
of the Fifth International Joint Conference on Artificial lntel/igence. Pittsburgh,
PA: Carnegie-Mellon University, pp. 833-842.
___ . 1982. "AM: An Artificial Intelligence Approach to Discovery in Mathematics
as Heuristic Search." In Knowledge-based Systems in Artificial Intelligence, eds.
Randall Davis and Douglas B. Lenat. New York: McGraw-Hill.
Lifschitz, Vladimir. 1985 "Computing Circumscription." Proceedings of the Fourth
International Joint Conference on Artificial Intelligence, pp. 121-127.
Lindsay, R.K., B.G. Buchanan, E.A. Feigenbaum, and J. Lederberg. 1981. Applications
of Artificiallntel/igence for Organic Chemistry: The DENDRAL Project. New York:
McGraw-Hill.
Llewellyn, Karl. 1960. The Common Law Tradition: Deciding Appeals. Boston: Little,
Brown.
Luttwak, Edward. 1979 [1967]. Coup d'Etat: A Practical Handbook. Cambridge, MA:
Harvard University Press.
MACSYMA Group. 1974. The MACSYMA Reference Manual. Cambridge, MA: Mas-
sachusetts Institute of Technology.
Mallery, John C. 1987. "Computing Strategic Language: Natural Language Models
of Belief and Intention." Paper prepared for the annual meeting of the International
Studies Association, Washington, D.C.
___ . 1988a. "Identifying Implicit Categories in Text Models: A Step in Com-
putational Hermeneutics." Paper presented at the Fourteenth World Congress of
the International Political Science Association, Washington, D.C.
___ . 1988b. "Thinking About Foreign Policy: Finding an Appropriate Role for
Artificially Intelligent Computers." Master's thesis. Cambridge, MA: MIT, De-
partment of Political Science, February 1988.
March, James G., and Johan P. Olsen. 1984. "The New lnstitutionalism: Organizational
Factors in Political Life." American Political Science Review 78: 734-7 49.
March, James G., and Herbert A. Simon, with the collaboration of Harold Guetzkow.
1958. Organizations. New York: John Wiley.
Martin, William A., and R.J. Fateman. 1971. "The MACSYMA System." Proceedings
of the Second Symposium on Symbolic and Algebraic Manipulation. Los Angeles,
pp. 59-75.
McCarthy, John. 1988. "Mathematical Logic in Artificial Intelligence." Daedalus 117,
1: 297-311.
McCarthy, John, and Patrick H. Hayes. 1969. "Some Philosophical Problems from
the Standpoint of Artificial Intelligence." In Machine Intelligence 4, eds. Bernard
Meltzer and Donald Michie. Edinburgh: Edinburgh University Press, 1969, pp.
463-501.
McCiosky, Donald N. 1985. The Rhetoric of Economics. Madison, Wl: University of
Wisconsin Press.
Mefford, Dwain. 1984. "Formulating Foreign Policy on the Basis of Historical Analogies:
An Application of Developments in Artificial Intelligence." Paper prepared for the
annual meeting of the International Studies Association, Atlanta.
___ . 1987a. "Analogical Reasoning and the Definition of the Situation." In New
Directions in the Comparative Study of Foreign Policy, eds. Charles F. Hermann,
Charles W. Kegley, Jr., and James N. Rosenau. London: Alien and Unwin, pp.
221-244.
- - - . 1987b. "The Policy Process and the Evolution of Argument." Paper prepared
for the Annual Meeting of the International Studies Association, Washington, D.C.
___ . 1987c. "The Cognitive Elements of a Theory of Foreign Policy." Presented
at the annual meeting of the American Political Science Association, Chicago.
- - - . 1988a. "The Acquisition of New Schema as a Function of Frustrated
Expectations and Surprise." Mershon Workshop on Change in International Politics.
Columbus, OH: Ohio State University, January 21.
- - - . 1988b. "Extracting Lessons from Experience: One of the Mechanisms in
a Memory-Based Model of Foreign Policy." Presented at the Mershon Seminar
in Foreign Policy and Decision Making, Columbus, OH.
92 Dwain Mefford
- - - . 1988c. "Women, Fire, and Artificial Intelligence," Paper prepared for the
conference "Words, Deeds and Social Things." Oxford University.
- - - . 1988d. "Modeling Conceptual Change in Political Institutions: The Case of
the Commerce Clause and White Slavery." Paper prepared for the annual meeting
of the International Studies Association, St. Louis, MO.
- - - . 1988e. "Cognitive Models of U.S. Interventions: The Function of Stories in
the Formulation of Foreign Policy." Paper presented at the combined Harvard-
MIT seminar, International Institutions and Cooperation, May 5, 1988.
- - - . 1988f. "An Alternative to Rational Choice: Narrative Reasoning and the
Structuring of Decisions." Paper presented at the Pacific Northwest International
Security Colloquium, University of Washington, May 29, 1988.
- - - . 1988g. The Analysis of Interdependent Decision-Making. Final Report,
National Science Foundation Grant #SES-8520307, July 1988.
- - - . 1988h. "Casuistry and Computer Science: Case-Based Reasoning in the
Realms of Moral Conduct, Law, and Politics." Paper presented at the triennial
meeting of the International Political Science Association, Washington, D.C.
___ . 1990. "Case-Based Reasoning, Legal Reasoning, and the Study of Politics."
Political Behavior 12, 2: 125-158.
___ . 1989b. "Learning in Soviet-American Relations: Constructing Abstract
Categories from Historical Events." Paper prepared for the annual meeting of the
International Studies Association, London.
___ . Forthcoming. Rationality and Reason in International Relations.
Michalski, Ryszard S. 1983 "A Theory and Methodology of Inductive Learning."
Artificial Intelligence 20: 111-161.
Miller, George A., E. Galanter, and Karl H. Pribram. 1960. Plans and the Structure
of Behavior. New York: Henry Holt, 1960.
Minsky, Marvin L. 1967. Computation: Finite and Infinite Machines. Englewood Cliffs,
NJ: Prentice-Hall.
Mitchell, Tom M., Richard M. Keller, and Smadar T. Kedar-Cabelli. 1986. "Explanation-
based Generalization: A Unifying View." Machine Learning 1, 1: 47-80.
Mooney, Raymond J. 1985. "Generalizing Explanations of Narratives into Schemata."
Proceedings of the Third International Machine Learning Workshop. Skytop, PA,
1985, pp. 126-128.
Mooney, Raymond J., and Scott W. Bennett. 1986. "A Domain-Independent Expla-
nation-Based Generalizer." Proceedings of Fifth National Conference on Artificial
Intelligence. Los Altos, CA: Morgan Kaufmann, pp. 551-555.
Mooney, Raymond J., and Gerald DeJong. 1985. "Learning Schemata for Natural
Language Processing." Proceedings of the Ninth International Joint Conference
on Artificial Intelligence. Los Altos, CA: Morgan Kaufmann.
Newell, Alien. 1980. "Physical Symbol Systems." Cognitive Science 4: 135-183.
___ . 1983a. "Cognitive Science and Artificial Intelligence." News Report. National
Academy of Sciences 33, 10(December): 3A-5A.
___ . 1983b. "The Heuristic of George Polya and Its Relation to Artificial
Intelligence." In Methods of Heuristics, eds. R. Groner, M. Groner, and W.F.
Bischof. Hillsdale, NJ: Lawrence Erlbaum.
___ . 1983c. "Intellectual Issues in the History of Artificial Intelligence." In The

Study of Information: Interdisciplinary Messages, eds. Fritz Machlup and U.
Mansfield. New York: John Wiley, pp. 187-227.
___ . 1989. "Putting lt All Together." In Complex Information Processing: The
Impact of Herbert A. Simon, eds. David Klar and Kenneth Kotovsky. Twenty-
first Carnegie-Mellon Symposium on Cognition. Hillsdale, NJ: Lawrence Erlbaum,
pp. 399-440.
Newell, Alien, J.C. Shaw, H.A. Simon. 1958. "Elements of a Theory of Human
Problem-Solving." Psychological Review 65: 151-66
Newell, Alien, and Herbert A. Simon. 1956. "The Logic Theory Machine: A Complex
Information Processing System." Transactions on Information Theory. Institute
of Radio Engineers, vol. IT-2, no. 3, pp. 61-79.
_ _ . 1963. "GPS, A Program that Simulates Human Thought." In Computers
and Thought, eds. E.A. Feigenbaum and J. Feldman. New York: McGraw-Hill,
pp. 279-293.
_ _ . 1972. Human Problem Solving. Englewood Cliffs, NJ: Prentice-Hall.
___ . 1976. "Computer Science as Empirical Inquiry: Symbols and Search."
Communications of the ACM 19, 3: 113-126.
Nilsson, Nils J. 1980. Principles of Artificial Intelligence. Palo Alto, CA: Tioga
Publishing Co.
Pednault, Edwin P.D. 1987. "Formulating Multiagent, Dynamic-World Problems in
the Classical Planning Framework." In Reasoning About Actions and Plans, eds.
Michael P. Georgeff and May L. Lansky. Los Altos, CA: Morgan Kaufmann, pp.
47-82.
Peirce, Charles Sanders. 1950. "The General Theory of Probable Inference." In The
Philosophy of Peirce: Selected Writings, ed. Justus Buchler. New York: Harcourt
Brace, pp. 190-217.
Phillips, Warren R. 1987. "Alternative Futures in the Middle East: The Results from
Three Simulations." In Artificial Intelligence and National Security. ed. Stephen
Cimbala. Lexington, MA: Lexington Books, pp. 161-186.
Polanyi, Michael. 1958. Personal Knowledge. Chicago: University of Chicago Press.
Polya, George. 1945. How To Solve it. Princeton, NJ: Princeton University Press.
___ . 1962. Mathematical Discovery. Vol. 1, On Understanding: Learning, and
Teaching Problem Solving. New York: John Wiley & Sons.
Post, Emil. 1943. "Formal Reductions of the General Combinatorial Problem." American
Journal of Mathematics 65: 197-268.
Radin, Max. 1933. "Case Law and Stare Decisis: Concerning Praejudizienrecht in
Amerika." Columbia Law Review 33, 2: 199-212.
Rapoport, Anatol. 1983. Mathematical Models in the Social and Behavioral Sciences.
New York: John Wiley & Sons.
Riesbeck, Christopher K. 1975. "Conceptual Analysis." In Conceptual Information
Processing, ed. Roger C. Schank. Amsterdam: North-Holland.
Riesbeck, Christopher K., and Roger C. Schank. 1989. Inside Case-Based Reasoning.
Rissland, Edwina L., and Kevin D. Ashley. 1986. "Hypotheticals as Heuristic Device."
Proceedings of the Fifth National Conference on Artificial Intelligence. Los Altos,
94 Dwain Mefford
Robinson, John Allan. 1965. "A Machine-oriented Logic Based on the Resolution
Principle." Journal of the Association for Computing Machinery 12: 24-41.
- - - . 1979 Logic: Form and Function. The Mechanization of Deductive Reason.
New York: North-Holland.
Rosenbloom, Paul S., and John E. Laird. 1986. "Mapping Explanation-Based Gen-
eralization onto SOAR." Proceedings of the Fifth National Conference on Artificial
Intelligence. Los Altos, CA: Morgan Kaufmann, pp. 561-567.
Sacerdoti, Earl D. 1974. "Planning in a Hierarchy of Abstraction Spaces." Artificial
Intelligence 5, 2: 115-135.
- - - . 1977. A Structure for Plans and Behavior. New York: North-Holland.
Schank, Roger C. 1972. "Conceptual Dependency." Cognitive Psychology 3: 552-
631.
- - - . 1975. Conceptual Information Processing. Amsterdam: North-Holland.
___ . 1982. Dynamic Memory: A Theory of Reminding and Learning in Computers
and People. New York: Cambridge University Press.
Schank, Roger C., and Robert Abelson. 1977. Scripts, Plans, Goals and Understanding:
An Inquiry into Human Knowledge Structures. Hillsdale, NJ: Lawrence Erlbaum.
Schank, Roger C., and Christopher K. Riesbeck, eds. 1981. Inside Computer Un-
derstanding: Five Programs Plus Minatures. Hillsdale, NJ: Lawrence Erlbaum.
Schrodt, Philip A. 1985. "Adaptive Precedent-Based Logic and Rational Choice: A
Comparison of Two Approaches to the Modeling of International Behavior." In
Dynamic Models of International Conflict, eds. Urs Luterbacher and Michael D.
Ward. Boulder, CO: Lynne Rienner Publishers, pp. 373-400.
Sergeev, Vicktor M., and P.B. Parshin. 1989. "Acquiring 'The Sense of Interdependence'
Through Crisis: An Essay in Cognitive Analysis of the Soviet Position on the
Caribbean Crisis of 1962." Presented at the Second Joint U.S. National Academy
of Sciences/ Academy of Sciences of the USSR Workshop on Models and Concepts
of Interdependence Between Nations, Tallin, Estonian SSR, USSR, January 4-8.
Shortliffe, Edward H. 1976. MYC/N: Computer-Based Medical Consultation. New York:
American Elsevier.
Simon, Herbert A. 1976. "Cognition and Social Behavior." In Cognition and Social
Behavior, eds. John S. Carroll and John W. Payne. Hillsdale, NJ: Lawrence Erlbaum,
pp. 253-267.
___ . 1980. "Cognitive Science: The Newest of the Artificial Sciences." Cognitive
Science 4: 33-46.
___ . 1985. "Human Nature in Politics: The Dialogue of Psychology with Political
Science." American Political Science Review 79, 2: 293-304.
Simon, Herbert A., and Michael Barenfeld. 1969 "Information Processing Analysis
of Perceptual Processes in Problem Solving." Psychological Review 76: 473-483.
Reprinted in Simon, Models of Thought. New Haven: Yale University Press, 1979,
pp. 363-372.
Simon, Herbert A., and Edward A. Feigenbaum. 1964. "An Information-Processing
Theory of Some Effects of Similarity, Familiarization, and Meaningfulness in Verbal
Learning." Journal of Verbal Learning and Verbal Behavior 3: 385-396.
Simon, Herbert A., and K. Gilmartin. 1973. "A Simulation of Memory for Chess
Positions." Cognitive Psychology 5: 29-46.
Stefik, M. 1981. "Planning with Constraints (MOLGEN: Part 1)." Artificial Intelligence
16, 2: 111-40.
Sternberg, Saul. 1963. "Stochastic Learning Theory." In Handbook of Mathematical
Psychology, vol. 2, ed. R. Duncan Luce, Robert R. Bush, and Eugene Galanter.
New York: Wiley, pp. 1-120.
Sycara, Katia P. 1985a. "Arguments of Persuasion in Labor Mediation." Proceedings
of the Ninth International Joint Conference on Artificial Intelligence. Los Altos,
___ . 1985b. "Persuasive Argumentation in Resolution of Collective Bargaining
Impasses." Proceedings of the Seventh Cognitive Science Society Conference.
Hillsdale, NJ: Lawrence Erlbaum Associates, pp. 356-360.
___ . 1988a. "Using Case-Based Reasoning for Plan Adaptation and Repair." In
Case-Based Reasoning, ed. Janet Kolodner. San Mateo, CA: Morgan Kaufmann.
pp. 425-433.
___ . 1988b. "Resolving Goal Conflicts via Negotiation." Proceedings of the
Seventh National Conference on Artificial Intelligence. San Mateo, CA: Morgan
Sylvan, Donald A., Ashok Goel, and B. Chandrasekaran. 1988. "Modeling Compiled
Structures in Case-Based Foreign Policy Decision Making." Presented at the Third
World Congress of the International Peace Science Society, College Park, Maryland.
Tambe, Milind, and Alien Newell. 1988. "Some Chunks Are Expensive." In Proceedings
of the Fifth International Conference on Machine Learning, San Mateo, CA: Morgan
Tanaka, Akihiko. 1984. "China, China Watching, and CHINA-WATCHER." In Foreign
Policy Decision Making: Perception, Cognition, and Artificial Intelligence, ed.
Donald A. Sylvan and Steve Chan. New York: Praeger.
Taylor, Charles Lewis, and David A. Jodice. 1983. World Handbook of Political and
Social Indicators. Vol. 2, Political Protest and Government Change. 3d edition.
New Haven: Yale University Press.
Thompson, William R. 1974. "Toward Explaining Arab Military Coups." Journal of
Political and Military Sociology 2: 237-50.
- - - . 1975. "Systemic Change and the Latin American Military Coup." Comparative
Military Studies 7: 441-59.
Thorson, Stuart, and Donald A. Sylvan. 1982. "Counterfactuals and the Cuban Missile
Crisis." International Studies Quarterly 26, 4: 539-571.
Van Melle, W. 1980. "A Domain-Independent System that Aids in Constructing
Knowledge-based Consultation Programs." Ph.D. dissertation, Stanford, Ca., Stan-
ford University.
Vosniadou, Stella, and Andrew Ortony, eds. 1989. Similarity and Analogical Reasoning.
New York: Cambridge University Press.
Waltz, David L. 1988. "The Prospects for Building Truly Intelligent Machines."
Daedalus 117, 1: 191-212.
Wendt, Alexander. 1987. "The Agent-Structure Problem." International Organization
41, 3: 335-370.
Winograd, Terry. 1972. Understanding Natural Language. New York: Academic Press.
- - - . 1983. Language as a Cognitive Process. Vol. 1. Syntax. Reading, MA:
Addison-Wesley.
96 Dwain Me/ford
Winston, Patrick H. 1970. "Learning Structural Descriptions from Examples." Report

No. TR-231. Cambridge, MA: Artificial Intelligence Laboratory, Massachusetts
Institute of Technology.
___ . 1975. "Learning Structural Descriptions from Examples." In The Psychology
of Computer Vision, ed. Patrick H. Winston. New York: McGraw-Hill. pp. 157-
209.
___ . 1979. "Learning by Understanding Analogies." Artificial Intelligence Labo-
ratory Memo 520. Cambridge, Mass. Artificial Intelligence Laboratory, Massa-
chusetts Institute of Technology, April 1979 (revised June 1979).
___ . 1980. "Learning and Reasoning by Analogy." Communications of the ACM
23, 12: 689-703.
___ . 1981. "Learning New Principles from Precedents and Exercises." Artificial
Intelligence 19: 321-350.
Winston, Patrick H., Thomas 0. Binford, Boris Katz, and Michael Lowry. 1983.
"Learning Physical Descriptions from Functional Definitions, Examples, and
Precedents." Proceedings of the Third National Conference on Artificial Intelligence.
Los Altos, CA: Morgan Kaufmann, pp. 433-439.
Winston, Patrick H., and Karen A. Prendergast, eds. 1984. The AI Business: The
Commericial Uses of Artificial Intelligence. Cambridge, MA: MIT Press.
Woodard, Calvin. 1987. "Common Law and Common-law Legal Systems." In En-
cyclopedia of the American Judicial System, ed. Robert J. Janoski. New York:
Scribner's, vol. 2, pp. 500-516.
4
Text Modeling for International Politics:
A Tourist's Guide to RELATUS 1
Hayward R. Alker, Jr., Gavan Duffy,
Roger Hurwitz, and John C. Mallery
And if this miracle [of human speech] is the key to the history of mankind, then
it is also the key to the history of society. . . . [Words] are capable of being
rays of light in a realm of darkness. . . . They are equally capable of being lethal
arrows . . . and even both at once!
-Vaclav Havel, 1989
Words, Deeds, Silences, and International Politics

Events in Eastern Europe in 1989 present an enormous challenge to
scholars and practitioners of international politics: How can we understand,
explain, project into the future, and beneficially respond to the revolutionary
changes that have taken place there? What does it mean to say that the
"Cold War" is "over"? What was the astonishing chemistry of connections
between Gorbachev's perestroika and the extraordinarily rapid and creative,
but mostly peaceful revolutions in the various smaller Eastern and Central
European countries, such as East Germany, Poland, Hungary, Czechoslovakia,
and Romania? How will "socialist" economic justice concerns of elites and
masses be reconciled with democratic "freedoms" and the "efficiency" of
market-oriented economics? (Hoffmann 1990)
Reflecting on these issues, Vaclav Havel, the new president of Czecho-
slovakia, has addressed the "mysterious link between words and peace
. . . the mysterious power of words in human history" (Havel 1989: 5).
Like many contemporary cognitive and social theorists, Havel sees words
both as intrinsically linked to human creativity and as essential for the
communication that makes economics and politics possible within and
among human societies. Words mediate or actually create "Man, Spirit, the
human soul, our self-awareness, our ability to generalize and think in concepts,
97
98 H. R. Alker, Jr., G. Duffy, R. Hurwitz, and J. C. Mallery
to perceive the world as the world (and not just as our locality) and . . .
our capacity for knowing that we will die" (Havel 1989: 5). Although they
can be liberating, words, speeches and texts are also ambiguous, ambivalent,
dangerous, and deceptive.
What about those who argue that, especially in international politics,
events or deeds speak louder than words? Where, in a perspective sensitive
to words and texts, are the pregnant or ominous silences? It takes but a
moment's thought to realize that they, too, do not escape the signifying,
mediating, constituting, disabling, and enabling power of symbolic, linguistic
communication. Artillery shells and tank treads may destroy life and property;
but hostile deeds are also, intrinsically, acts of political communication, pain-
bringing physical extensions of a contest of political wills, intelligible and
purposive acts for which their agents are held responsible. The existence
of empty, nearly soundproof torture rooms speaks volumes about the bases
of legitimacy of regimes that use them. Unexamined habits of compliance
with once controversial international standards of economic conduct are just
as politically relevant phenomena as the words and texts used reflectively
to constitute, plan, justify, delineate, interpret, or modify such acts, deeds,
standards, practices, or institutions.
RELATUS: An Environment for the

Scientific Study of Texts
An Outline of the Present Chapter

In the hope of eventually making such word-infused phenomena as
accessible to rigorous scientific analysis as are any other so characterized,
whole disciplines of language, text, and discourse study have emerged. As
a computationally oriented contribution to these disciplines, we informally
introduce here the operation, uses, potentials, and limitations of the RELATUS
system. 2 Together, its routines 3 provide a computationally manipulable natural
language environment for what we shall describe as "text modeling" and
"text analysis." As our major task we shall motivate and describe RELATUS's
basic procedures for replicably creating what we call "text models" from
"literal and explicit" textual inputs. RELATUS is designed to simulate the
role in interpretation of the "prejudicial horizons" of background knowledge,
and the intentional, instrumental, and expressive, situation-constituting ca-
pacities of particular social actors (including individual social scientists). We
shall demonstrate in several ways the role of cultural and belief-system-
specific backgrounds in creating such text models (see Alker 1975; Alker,
Lehnert, and Schneider 1985; Duffy and Mallery 1986; Alker 1988; Duffy
1987; and Mallery, Hurwitz, and Duffy 1987 for our earlier attempts to
synthesize parts of the hermeneutic traditions and computational linguistics).
Although its utility for replicable text modeling and analysis in many
other domains and disciplines will be suggested, we shall focus here on
RELATUS's potential contributions to the study of international/world politics.
Indeed, the single example worked through here will come from a program
of empirical research on international conflict management initiated by Ernst
B. Haas (Haas 1967-1968; Butterworth, with Scranton 1976; Alker, Bennett,
and Mefford 1980; Alker and Sherman 1982; Haas 1986; Sherman 1987!
1988). Because of its focus on the sociohistorical embeddedness of inter-
national institutions and the importance of interpretive practices in their
development, this research may be identified with the "reflectivist" side of
contemporary arguments with game theoretic "rationalists" about how best
to account for the existence, performance, and developmental possibilities
of international institutions (Keohane 1988; Alker and Ashley, in preparation).
Noting for the sake of generality that RELATUS's input texts may be
reformulated from a variety of original and secondary sources, such as news
reports, archival documents, experimental protocols, and even textually
reconstituted statistical data, in the present paper we shall illustrate RELATUS-
based formal representations derived from the Haas-instigated research
program. Our primary illustration is a theoretically structured historical
narrative of the 1956 Soviet intervention into Hungary. Written by Robert
Butterworth (a Haas Ph.D. student), with help from Margaret Scranton, this
two-page account, repeated in Table 4.1 below, links these developments
causally to the de-Stalinization dynamics of the Soviet bloc; it also recounts
the key substantive claim of the limited but pro-Western response of the
United Nations General Assembly at that time. lt is also possible to transform
conventional "quantitative" data into RELATUS-parsable sentence forms
(Mallery 1988b) capable of being compared, contrasted, and referentially
integrated with wholly textual accounts. 4 Frank Sherman's SHERFACS re-
lational data base allows many such synthetic possibilities, to be discussed
elsewhere (Sherman, 1988). 5
We shall leave to a separate, lengthier exposition in this volume Gavan
Duffy's "time base sorting" and John Mallery's "semantic content analysis."
In conclusion, we shall return to the philosophical and methodological
challenge for political science posed by creative politics, such as the world
saw happen in 1989 in Eastern Europe. RELATUS-based research is a
methodological effort to respond to the historically and interpretively reflective
aspects of such creativity, often neglected by other social scientific research
approaches. lt uses parsability and constraint posting to make possible
efficient, domain-independent, context-sensitive modeling and analysis. Its
100
Table 4.1: Butterworth-Scranton's "Narrative Data" on the Hungarian Intervention

(Dispute #118)
Parties: USSR v. Hungary
Agents: UN
The Hungarian crisis, like so many of the East European "disturbances" at
this time, stemmed from the ideological and power struggles which resulted from
the Soviet-led deStalinization process (see case #117: Polish October, 1956). Its
climactic phase began with the repercussions of Georgi Malenkov's fall from
power in the USSR in 1955 for "weakening the Party in favor of the state." The
Hungarian adaptation of Malenkov's "new Course" two years earlier had led to
the development of a sharp and well-defined ideological dispute between the
Hungarian Premier, lmre Nagy, and the super-Stalinist General Secretary of the
Hungarian Communist Party, Matyas Rakosi. Malenkov's fall cut vital support
for Nagy and enabled Rakosi to effectuate Nagy's expulsion from the Politburo,
the Cental Committee, the Premiership, and eventually even the Party itself.
However, according to Brzezinski, the earlier Nagy period (1953-55) was particularly
significant in that it created a conception of an alternative economic and political
policy within the existing framework.
The Nagy-Rakosi debate continued, spilling from the elite to the mass level,
with the formation of "intellectual" Petofi Circles throughout the country. The
Nagy program, infused with popular support, had by now evolved into "national
communism." DeStalinization in the USSR, inaugurated with Khrushchev's shift
from a hard to a soft line, had now cut the support that Rakosi badly needed.
His reaction to the Poznan riots in Poland (see case #117) and to the growing
restlessness within Hungary was to attempt to arrest Nagy and his associates
and to repress all dissident organizations. His drastic plans met stiff resistance
within the Central Committee and resulted instead in his own resignation. A
Soviet-sponsored compromise candidate, Gero, who was actually a neo-Stalinist
associate of Rakosi, was chosen as his replacement.
This choice satisfied virtually no one. With the Soviet acceptance of Gomulka
in Poland as Party chief, the now politicized masses, viewing this development
as a legitimization of national communism, took to the streets to demand the
resignation of Gero and his replacement by Nagy. When Gero's last element of
support, the secret police, fired upon pro-Nagy demonstrators, the country was
thrown into utter chaos. Nagy was reelected as Premier by the Hungarian Cental
Committee, and a few days later Gero was placed as First Secretary by Janos
Kadar. In the continued violence the Party had virtually disintegrated, and what
government that did exist seemed hopelessly divided. Nagy appealed for order,
promising to follow a "Hungarian road to socialism," while Kadar appealed for
vigilance against "counterrevolutionaries." One member of the Hungarian Politburo
called for direct Soviet military aid.
Soviet troops within Hungary entered the sproadic fighting which was raging
throughout the country. On 28 October Nagy negotiated a cease-fire, and Soviet
troops began to pull out of Budapest on the following day. Bilateral negotiations
between representatives of the Hungarian and Soviet parties broke down, however,
as the involvement of Soviet troops had already radicalized the demands of the
rebels and pushed Nagy over the brink of acceptability. Nagy eventually denounced
the Soviet intervention, announced the setting up of a multi-party political system,
began negotiations for Hungary's withdrawal from the Warsaw Pact, and officially
proclaimed Hungary's neutrality.
(continues)
On 27 October France, the UK, and the US requested UN Security Council

discussion of the situation in Hungary, claiming that the rights of the Hungarian
people were being repressed by foreign military forces. The following day Hungary
protested consideration by the Council of events within Hungary's domestic
jurisdiction. However, on 1 November Premier Nagy informed the UN Secretary-
General that additional Soviet troops had entered Hungary, and on 2 November
he requested that the Council instruct the Hungarian and Soviet governments
to negotiate their differences. Soviet opposition deadlocked the Council over the
question, and so the issue moved (under the 1950 "Uniting for Peace" procedure)
to the General Assembly. On 4 November the Assembly adopted a resolution
calling on the USSR to withdraw all its forces from Hungary without delay.
However, by this time the second Soviet intervention successfully put down the
Hungarian revolt, and the new Hungarian government of Janos Kadar informed
the Secretary-General that all communications from Nagy were invalid and it
then reiterated Hungary's objection to discussion of the situation in the UN.
The Assembly persisted, however, even though it was clear that Soviet policy
would not be deflected and that the uprising had been crushed. In a series of
resolutions the Assembly called upon the USSR to undertake a variety of actions,
called upon the UN Secretary-General to undertake initiatives, and established
a UN Special Committee on the Problem of Hungary on 10 January 1957. The
Committee's investigations (although it was never able to visit Hungary) led it
to conclude that the revolt had been spontaneous and nationalistic, had started
peacefully but had been escalated by the violence of political police firing upon
people, and had been legally and properly considered by the UN. This report
was endorsed by the Assembly, which also denounced (12 December 1958) the
subsequent execution of lmre Nagy and three of his associates. The Assembly
continued to pass resolutions in this vein through 1962.
Refs: Everyman's {United Nations}., 192-193.

Zbigniew Brzezinski, The Soviet Bloc: Unity and Conflict, Cambridge, Mass.:
Harvard University Press, 1967, 210-238.
Source: Butterworth, with Scranton, 1976, pp. 217-219. Reprinted with per-
mission of the publisher, the University Canter for International Studies at the
University of Pittsburgh.
text models are designed to be consistent with a "lexical-interpretive"se-

mantics and a word-sensitive "evolutionary constructivist" epistemology.
A Preview Outline of RELATUS Text Processing

Each sentence of a simplified input text must be computationally analyzed,
or "parsed," into a representation structure before it can be integrated into
a text model. After a brief review of Duffy's efficient, transformational parser,
we shall show how Mallery's sentential constraint posting (i.e., "attaching")
process distills the parser's clausal and phrasal analyses into semantic
constraints. During the constraint-posting phase, grammatical relations from
deep structure and quantified noun phrases from surface structure are
expressed as a "reference specification" -a "logical form" that carries the

sentence's propositional content. Next a reference system finds or creates
knowledge structures corresponding to the specification. The sequential
repetition of this process for the sentences of background and input texts
yields a referentially integrated "text model," or semantic network.
Put into a formula, the basic sequence of RELATUS text modeling is:
(Syntactic representations and processes + semantic representations and

processes) applied to T ("literal and explicit" input Text, including new
background information) - referentially integrated semantic network
text model of T.
In order to turn input text (or texts) into RELATUS "text models," one
must rewrite them into "literal and explicit" English6 and supplement each
simplified, reformulated text with background texts containing culturally
general or individually specific knowledge and beliefs. These, too, must be
referentially integrated into the belief-system-specific "text models." RE-
LATUS's associated theory of text linguistics provides the "data theory" of
RELATUS "data making," and the resulting text models are the "data"
whose reliability and validity a careful social scientist should want to assess.
These concerns call for serious, additional work identifying appropriate
cultural and belief-system-specific background knowledge. 7 As with any other
natural language processor, RELATUS forces its users to be explicit about
the conceptual and interpretive presuppositions brought to textual analysis.
Preparing for Belief-System-Specific Text Modeling
The Text on the Hungarian Case

The 800-word Butterworth-Scranton text on the "Hungarian Intervention,
1956" is given in Table 4.1. We shall discuss its contents from several
perspectives, the first being Alker and Sherman's reactions to previous work
by Haas and Butterworth.
Alker and Sherman's Research Concerns

Working in the Haas-Butterworth tradition but dissatisfied both with its
restricted data bases and modes of processual assessment (using only cross-
tabulations of cases actually considered by international conflict management
agents, as obtained from a rather limited set of sources), Alker and Sherman
were especially concerned to create richer data sets for comparative analysis,
narrative modeling, and simulational studies of collective security conflict
management practices (Aiker and Sherman 1982; Alker, Bennett, and Mefford
1980). Sherman was especially interested in the agenda processes wherein

conflicts, disputes, or quarrels were originally conceptualized as being of
international concern. Extensive further data-gathering activity is reported
in his dissertation (Sherman 1987), which became the basis for the SHERFACS
data set (including available conflict narratives written from multiple per-
spectives), which superceded and subsumed Alker and Sherman's earlier
data set (Sherman 1988).
Although Alker's and Sherman's efforts eventually produced on the order
of 150 more punched cards per case than either Haas or Butterworth's
single-card codings, they were still dissatisfied with the way such codings
chopped up their textual sources, burying certain controversial biases. For
example, "post mortem" discussions of the Hungarian intervention have
raised important, controversial questions about the conflict-exacerbating role
of the Voice of America. Butterworth and Scranton do not mention its calls
for a popular uprising in Hungary, with intimations that U.S. aid might be
forthcoming; neither do they mention the structural effects of the Cold War
bipolar conflict on Soviet or U.S. actions. The motivational basis of Nagy's
courageous but tragic proclamation of (socialist?) neutrality is omitted. Such
policy argumentation, tragically like the Melians' plea for neutrality in
Thucydides's famous history, needs to be recognized (Aiker 1988).
Consider also Butterworth-Scranton's presentation as "data" of a Brze-
zinski-inspired structural interpretation of the causes of the "disturbance,"
as a contradictory consequence of Soviet-led de-Stalinization. This is pro-
vocative, but controversial. It treats Eastern bloc "disturbances" as almost
necessary consequences of bloc-internal contradictions, without mentioning
either the Cold War or U.S. actions or the possibility of similar consequences
of "unequal alliances" in the West. Such implications require critical,
comparative examination, especially if one is interested in building larger,
less ideological, testable, and critically confirmed generalizations about the
international order in Europe and beyond. As illustrated in a polimetric
discussion of the Congo crisis (Aiker 1975), multiple perspectives on
international issues, with the diplomatic processes of response thereto, could
be constructively analyzed, contrasted, and synthesized using rationally
structured narrative accounts.
Our Early Interests in Narrative Models

Having been an appreciative student and colleague of Robert Abelson at
Yale, Alker used Schank-Abelson conceptual dependency formalisms (in
Alker 1975; see Schank and Abelson 1977 for a later version) to formalize
a critical Third World narrative account of the "old" international economic
order. He regularly taught his "complex models" graduate students about
Abelson's two decades of work modeling ideological belief systems (see
104
Table 4.2 (A and B): The First Two Paragraphs of Butterworth's Hungarian
Intervention Text in Literal and Explicit Form, Plus Directly Relevant Background
Text, Both from an 8-page Text and Successfully Reentered into a RELATUS
Text Model, 4/16/1987
A. THE LITERAL AND EXPLICIT TEXT
The Hungarian intervention is an international crisis.

The Hungarian intervention involves Hungary, the USSR, Poland, France, the
UK, the USA, the UN-Security-Council, Georgi-Malenkov, lmre-Nagy, Matyas-
Rakosi, Khrushchev, Janos-Kadar, Hungary's masses, Hungary's elite, Gomulka,
and Gero.
- - Eastern Europe destalinized because the USSR destalinized.
Hungary destalinized because eastern Europe destalinized and because Hungary
is in eastern Europe.
The USSR's destalinization was in the USSR.
Eastern Europe's destalinization was in eastern Europe.
Hungary's destalinization caused some ideological struggles.
The ideological struggles were in Hungary.
The ideological struggles caused some power struggles.
The power struggles were in Hungary.
Georgi-Malenkov's New-Coursse (sic) was a political program.
Georgi-Malenkov's New-Course was the USSR's top-level policy because Georgi-
Malenkov was the USSR's premier and because he wanted the New-Course to
be the USSR's top-level policy.
The New-Course was not stalinist because the New-Course was the USSR's
destalinization.
Georgi-Malenkov was able to begin the USSR's destalinization because he was
the USSR's premier and because he was not a stalinist.
Georgi Malenkov was not a stalinist because Georgi-Malenkov's New-Course
was not stalinist.
Georgi-Malenkov fell from political power in the USSR because he attempted to
weaken the USSR's communist party and because the New-Course was not
stalinist.
Georgi-Malenkov attempted to weaken the USSR's communist party because he
wanted to strengthen the USSR's bureaucracy.
lmre-Nagy was a Hungarian premier.
lmre-Nagy was a Hungarian politburo member. lmre-Nagy was a Hungarian central
committee member. lmre-Nagy was a Hungarian communist party member.
Matyas-Rakosi was a general secretary of Hungary's communist party. Matyas-
Rakosi was a stalinist.
lmre-Nagy's political program was Hungary's destalinization because lmre-Nagy's
political program was similar to the New-Course.
Hungary's top-level policy was similar to the New-Course because lmre-Nagy's
political program was Hungary's top-level policy and because lmre-Nagy's political
program was not stalinist.
lmre-Nagy's political program caused an ideological dispute to occur.
Matyas-Rakosi and lmre-Nagy composed the major disputants of the ideological
dispute.
The USSR had supported lmre-Nagy.
lmre-Nagy's support by the USSR became low because Georgi-Malenkov fell
from political power in the USSR.
lmre-Nagy's political power became low because lmre-Nagy's support by the
USSR became low.
(continues)
105
Matyass(sic)-Rakosi's political power became high because lmre-Nagy's support

by the USSR became low.
Matyas-Rakosi was able to influence Hungary's communist party because Matyas-
Rakosi's power was high and because Matyas-Rakosi wanted to influence
Hungary's communist party.
Hungary's communist party forced lmre-Nagy to resign his political office beacuse
Matyas-Rakosi wanted lmre-Nagy to resign his political office and because Matyas-
Rakosi influenced Hungary's communist party.
The USSR supported Matyass(sic)-Rakosi.
The ideological dispute was initially an elite dispute.

The ideological dispute became a mass dispute because the ideological dispute
became widespread.
lmre-Nagy's political program acquired popular support because the masses
liked lmre-Nagy's political program.
The masses liked lmre-Nagy because the masses liked lmre-Nagy's political
program.
lmre-Nagy became popular because his political program acquired popular
support.
lmre-Nagy's political program evolved into national Communism.
The USSR destalinized again because Krushchev (sic) ascended to political
power in the USSR.
Krushchev (sic) initiated the USSR's destalinization in the USSR because he
shifted from a hard line to a soft line.
Matyas-Rakosi's support by the USSR became low because Krushchev (sic)
initiated the Soviet destalinization and because Matyas-Rakosi was a stalinist.
The Poznan riots and some Hungarian unrest caused Matyas-Rakosi to fear that
he would fall from power.
Matyas-Rakosi attempted to arrest lmre-Nagy because he feared that he would
fall from political power.
Matyas-Rakosi also atempted (sic) to arrest lmre-Nagy's associates because the
Poznan riots and the hungarian (sic) unrest caused Matyas-Rakosi to fear that
he would fall from political power.
Matyas-Rakosi also attempted to repress all dissident organizations in Hungary
because Matyas-Roakosi feared that he would fall from political power.
The Hungarian central committee strongly resisted Matyas-Rakosi's drastic plans.
Matyas-Rakosi resigned his political office because the Hungarian central com-
mittee forced him to resign his political office.
The central committee forced Matyas-Rakosi to resign because it strongly resisted
Matyas-Rakosi's drastic plans.
Gero was similar to Matyas-Rakosi because Gero was Matyas-Rakosi's associate.
Gero was similar to Matyas-Rakosi because Gero was a neo-stalinist and because
Matyas-Rakosi was a stalinist and because all neo-stalinists is (sic) similar to
all stalinists.
The USSR sponsored Gero for political office because the USSR wanted Gero
to be Hungary's premier.
The USSR wanted Gero to be Hungary's premier because Gero was a compromise
candidate and because the USSR did not like lmre-Nagy and because Gero was
similar to Matyas-Rakosi.
Gero became Hungary's premier because the Hungarian central committee chose
Gero as Matyas-Rakosi's replacement.
The central committee chose Gero as Matyas-Rakosi's replacement because the
USSR sponsored Gero as Hungary's premier and because the USSR was able
to influence the central committee.
Carbonell 1982 for an important milestone) and Abelson's unpublished,

"generative" models of fairy tales (see Alker 1987, first drafted in 1975).
He also taught the social science relevant work of MIT computer science
faculty and encouraged his technically capable students to study with those
faculty. Around 1981, Alker proposed to Duffy and Mallery that they try
computationally to model some of the 8utterworth-Scranton narrative texts,
a proposal Patrick Winston also made to Mallery. All knew of a certain
convergence between Vale-style narrative modeling, Alker's simulational
studies of precedential reasoning (Aiker, 8ennett, and Mefford 1980, a rewrite
of a successful National Science Foundation proposal written in 1977), lthiel
Pool's interest in applying artificial intelligence to content analysis, and
Winston's work on analogical reasoning (Winston 1980, 1981; Katz and
Winston 1982).
RELATUS was born out of discussions in 1982 and 1983 between Duffy
and Mallery about an alternative to the Winston-Katz ANALOG system,
whose broader coverage and space efficiency (minimal garbage production)
would better conform to their polimetric needs. As Duffy's and Mallery's
chapters in this volume will further detail, RELATUS now consists of many
subsystems, early among which were a parser and categorical disambiguator
(Duffy 1987), a poster of syntactic constraints useful for establishing eo-
references (Duffy and Mallery 1984; Mallery 1985; Mallery and Duffy 1990,
draft available since 1986), coreference and knowledge representation systems
(Mallery 1990, draft available in 1985), and many inversion, question-
answering, and interface routines written by Mallery for the Symbolics LISP
machine.
Inputting a Text
Oriented toward an as yet uncompleted comparative study of international
interventionist practices, RELATUS reached a milestone with the text given
in Table 4.2 (Mallery, 1987). In 1984, and again in 1987, its full, dense,
eight-page version was successfully transformed into RELATUS's semantic
network representations.8 Table 4.2A contains successfully parsed/repre-
sented "literal and explicit" RELATUS input text corresponding to the first
two paragraphs of the original 8utterworth narrative. Table 4.28 gives about
one-third of the background assumptions incorporated in the relatively
internationalist Cold War belief system we see this text to represent; only
terms or class hierarchies directly appearing in the simplified text in Table
4.2A-based on the first two paragraphs of the original document in Table
4.1-have been included. Some typographical errors have been included to
give a more realistic sense of these experimental computer runs. Looking
in more detail at Mallery's background information in Table 4.28, we see
how the "conflicts" and "struggles" and "fighting" of the earlier 8utterworth
107
Table 4.2 (continued)

B. DIRECTLY RELEVANT BACKGROUND BELIEFS (PLUS ;;COMMENTS)
Intervention is conflict. Disputes are conflicts. All struggles are conflicts.

A revolt is a conflict. All riots are revolts.
The Poznan riots were a dispute. Riots are unrest. (;;Omit?)
All elites are social groups. All masses are social groups. ;;make elite a noun
A general secretary is a political leader. A premier is a political leader.
A member of any communist party is a political leader. A member of any central
committee is a political leader. A politburo member is a political leader.
A premier is a political office-holder. A general secretary is a political office-
holder. A politburo member is a political office-holder. A central committee member
is a political office-holder. A member of any communist partyis a political office-
holder.
;;Should be parsed when class propagation is in.
;;A general secretary os a country's communist party is a general secretary, etc.
All communist parties are political parties.
A neo-stalinist is a stalinist.
;;A destalinization is not a stlinization. Parser can't handle this. To destalinize is
not to stalinize.
All neo-stalinists are similar to all stalinists because neo-stalinists are stalinists.
Every top-level policy is a political program.
;;Need to attach this to the verbal form. Looks screwy. A destalinization is a
political program.
A communist party is a social organization. Scial (sic) organizations are living
systems.
The USSR's communist party is a communist party. Hungary's communist party
is a communist party.
Interventions are processes. A conflict is a process. All destalinizations are
processes.
lmre-Nagy was a political leader. Hruschev (sic) was a political leader. Georgi-
Malenkov was a political leader. Matyas-Rakosi was a political leader. Gero was
a Hungarian political leader.
The USSR is a superpower. Poland is a lesser power. Hungary is a lesser power.
;;Should indicate being physically inside.
Poland is in eastern Europe. Hungary is in eastern Europe.
The USSR is a communist country. Poland is a communist country.
Hungary is a Soviet satellite. Poland is a Soviet satelliete.
The Poznan riots were in Poznan. Poznan is in Poliand (sic).
Gomulka was Poland's premier.
National communism is nationalistic.
Source: Mallery 1987.
paragraphs are linked to the "revolt," whose motivation is identified as

"spontaneous and nationalistic" near the end of the accounts in Table 4.1
(and in an extended version of Table 4.2A not included here). RELATUS
allows the preservation and interpretive investigation of such meaning
relationships. It thus shares an informational ideal with, but not the lin-
guistically impoverished formalisms of, contemporary political statistics.9
On the other hand, Table 4.2C goes beyond such inputs to report
referentially integrated details from a visually inspectable RELATUS belief
system. They are limited to the category "leader," its subcategories, and
their instances, as interconnected in the postwar conflict belief system
resulting from combining full versions of Tables 4.2A and 4.28. 10 We present
this additional example here in order to give a better sense of what RELATUS
background inputs contain.
A careful comparison of leader subclasses (with distinguishing postscripted
indexing numbers) shows the class hierarchy information in Table 4.28 has
helped RELATUS routines successfully group Soviet (postscripted -3),
Polish (- 4), and Hungarian (- 2) leader groups; premiers (-1) are an
overlapping group. The same Gero (indexed as -1) is both a LEADER -2
of Hungary and a (NEO-) STALINIST. The non-postscripted categories-
premier, leader, Stalinist in Table 4.2C-have relationships (as leaders,
premiers, or forms of mammal being) explicitly stated in the full text versions
of Table 4.2A and 4.28; the reader is encouraged to check through the
first parts of the table (and Table 4.1) to see how well this belief system
excerpt portrays meaningful and valid interconnections among these cate-
gories and their instances. The reader is also requested to explore what
recent developments suggest about the recategorizations appropriate for such
hierarchies: Should the secretary of the Communist party be considered
the same kind of "leader" as once would have been assumed? Is DubCek,
if he were to be included here, still to be considered a "socialist"-one of
those words whose meaning has changed, according to Have!?
Depending on the quality, novelty, and historical sensitivity of background
inputs, computerizing two or three pages of textual information for RELATUS
takes upwards of several hours for the skilled analyst, although the time
involved will decline as grammatical and referential coverage increase. 11
Constructing and Analyzing a Text Model

We shall now illustrate what RELATUS does with such inputs. Concretely,
we shall show how RELATUS helps build a text model starting with a "de-
Stalinization" sentence marked in Table 4.2A; to visualize much larger text
models than book publications allow, map-size visual aids, video cameras,
or computer-scrolling technologies are necessary. Although "text modeling"
in this chapter's title is meant to include many such subsequent steps, one
Table 4.2C: An Indexed Leader Class Hierarchy (derived from the parsed, referential integration of full text versions of Table 4.2A and
4.28)
{. BE-./9 J (PRRTICULRR-718 ) - ( PREftiER-1) (BE-././ 7 )-( GOftULKR-f)

( PRRTICULRR-884 ) - ( PREftiER-2) (BE-./ 3./ ) - ( CEORCI -ftRLEHKOII-1)
( PRRTICULRR-925 ) - ( PREftiER-3 ) (BE-./ 33 )-( STRLIHIST )

1.{ BE-28 -(PREIIIER)
( PRRTICULRR-984 ) - ( PREftiER-4) ( BE-./ 36 )-( IftRE -HRCV-1 )
',{ BE-2.1 , ... ----- .
( PRRTICULRR-1428 ) - ( PREftiER-5) (BE -.16./ ) - ( JRHOS-KRDRR-1 )
( LERDER-1 J K: BE-22 ,-(nEnBER-2)
(BE-./53 )-(CER0-1)
1
\( BE-23 - (, ---
n~;nuER-3) I( PRRTICULRR-1678 )-(PREniER-6 )----<(
(BE-./68)-(InRE-HRCV-1)
(BE-8-1) (IftRE-HRCV-1)
( BE-87) (nRTYRS-RRKOSI-1)
I'( BE-82)
~ (BE-88) (GER0-1)
(BE -89) ( JRHOS-KRDRR-1 )

I---- -- -- -- - '
('-8;:E;:-:::::;8:::;5~j l ~HKUI>!;HI;V-1 )
~ BE-83), LERDER-3
~ ,...
(BE -86) ( GEORGI -nRLEHKOII-1 )
.....
~
110 H. R. A/ker, Jr., G. Duffy, R. Hurwitz, and J. C. Mallery
could also describe our beginning exploration of the causal context of these
two sentences in the larger Soviet intervention in Hungary text model as
an elementary "text analysis" (de Beaugrande and Dressier 1981).
RELATUS achieves domain-independent natural language processing with
a bottom-up strategy that builds text models from syntactic analysis, logical
form construction, and an indexically complete reference system. Not only
may the researcher use it to study, compare, or contrast possibilities for
consensus formation among persons, groups, or governments with distinctive
belief systems, he or she may also study very different domains as well,
with the same natural language processing system augmented by appropriate
domain-specific dictionary additions and background beliefs or knowledge.
Parsing a De-Stalinization Sentence

Duffy's syntactic parser can parse many richly structured sentences,
involving a variety of embedded clauses. It uses a large dictionary that users
can easily extend as required by the texts they wish to model. The parser
disambiguates different parts of speech for particular words (Duffy 1986)
and recognizes semantically or pragmatically ambiguous clausal and phrasal
attachments, prompting the user to resolve them. The user may optionally
appraise parsing results during various stages of this syntactic process.
Given an English sentence (or text stream), a preparser first preprocesses
this string, returning a list with expanded contractions and annotations
marking capitalization, punctuation, and possessives. The parser then creates
a sentence object, such as that for the sentence (from Table 4.2A) in Figure
4. 1. The sentence object then queries GNOSCERE, the knowledge repre-
sentation system, for the current deictic context to resolve indexical pronouns
such as "you," "here," and "now." (See Levinson 1983 and the discussion
of conversational, object-oriented programming in Alker 1986 and Steele
1990, for more background on these technical matters.) Then control is
passed to some word categorization disambiguation procedures-which
distinguish, for example, "conflict" as a noun from "conflict" as a verb.
Clausal complements and other constituent boundary markers are inserted
where elided in the text, to simplify subsequent clausal and phrasal boundary
detection routines.
The 1989 version of Duffy's parser does phrase structure analysis-such
as that for the ADVERBIAL-1 "because" phrase of Figure 4.1-before it
attempts clausal analysis, facilitating detection of ambiguous attachments.
Unambiguous prepositional identification and intrasentential "anaphora res-
olution" procedures (binding each "this" and "that" to the appropriate
object) and intrasentential noun phrase coreferents (across different clauses)
are also identified by the processing at this stage. After these steps have
Figure 4.1: A RELATUS-Parsed Sentence from Table 4.2A

(sentence numbers do not correspond to original ones)
SENTENCE-274
Eastern Europe destalinlzed because the
USSR destalinlzed.
USSR PERFECT
Proper Noun Aspect
created a surface structure for a sentence, its clauses and phrases, the
parser applies grammatical transformations, such as the complementing and
tensing indicated in Figure 4. 1, to produce a declaratively canonical deep-
structure representation.
Figure 4.2 shows Mallery's indexed, referentially specified, constraint-
posted logical form for the previously parsed sentence. Note how the
constraints come out of the grammatical distinctions identified by the parser,
such as qualitative and quantitative implications about the kind and number
of subjects intended, and relational constraints, such as implied by the tense,
aspect, and varying truth-assertive character of verbs. The top one, for
example, states that the EUROPE-2 object in question, which stands in a
i\j
Figure 4.2: The Referential Specifications of Figure 4.1 with Preferential and Relational Constraints Posted on Them and Some References
Resolved
&UIJ[CI·RlUilroii) l .......... )
.[rf:c:;;o:;;n:;stt;:r•w•n;;;,;;.;-))-----1~ PIDPU-MOUft-P) Constr•in<s -@!)
- Constretnt-List
R•f-Sp•~
DfSTALUil£·1
PftSURJ£CT -IHRIIOM
HM-T£1'15L-IJI
PftOIJ£CT -IELAIJOM
C~·l
subject relation to the main de-Stalinize verb in the sentence, is itself

"quantificationally" constrained to be an individual proper noun. This textually
indexed EUROPE-2 is itself constrained by a single modifier, that it truly
"has the quality (HQ)" or value "EASTERN." The constraints are inspected
using the "REF-SPEC Examiner" RELATUS program. The "P(referred)" and
"P(referred)M(andatory)" prefixes of the grammatical subject and object
relations in both the REF-SPEC graphs of Figure 4.2 refer to properties
that are mandatory or preferred in the automated search for coreferences
to the same subjects, objects, or verbs.
Similar figures could also be presented for each of the literal and explicit
sentences of the full version of Table 4.28. Especially interesting complications
are suggested by the sentence: "The USSR's de-Stalinization was in the
USSR," to which we shall return in discussing Figure 4.3 below.
The full text model corresponding to the Hungary crisis narrative and
its eight dense pages of background text is a semantic network occupying
approximately half a megabyte of computer memory; hence it can be only
visualized partially at any one time. The first structure in Figure 4.3 and
the left half of the second one are text model parts-a belief system's
sentence forms-corresponding to the two sentences we have just discussed.
Generally read from top to bottom and right to left, and from subjects
through relations to their objects, these elements are given abstractively
simplified representations by RELATUS's belief system examiner. Italicized
CAUSE and PARTICULAR relations should, however, be read left to right.
Notice also that the verb nominalization of the second de-Stalinization
sentence referred to above has been represented in Figure 4.3 as implying
a verb form, DESTALINIZE-2, whose post-scripted "-2" identifier correctly
coreferences the USSR's de-Stalinization, as previously discussed in a different
sentence, viz. that represented in Figure 4. 1.
Interrelationships among three "DESTALINIZE" processes and three
grammatically tensed "CAUSE" relations are shown incrementally in Figure
4.3. What a new kind of statistically "overidentified" causal model we now
have: one where ontologies among forms of being and class hierarchies
among subjects and objects are also suggested, time relationships include
those suggested by natural language verb tenses, and "becauses" include
both intentional accounts and naturalistic causes as well!
Going well beyond early, simple identificational question-answering routines,
the only computational technique presently available for analyzing the
meanings of RELATUS-generated, qualitative text models is semantic content
analysis, as presented in Mallery's chapter in the present volume. The
semantic content analysis facilities allow researchers to formulate, collect,
store, and apply recognizers for knowledge structures. Once induced from
English sentences, recognizers can also label instances of the categories
they recognize. A researcher may inspect recognitions, like any other
.to.
-
Figure 4.3: A Partial RELATUS Text Model of Some Causal Interrelations Among De-Stalinization Processes in Belief System Corresponding
to the Complete Version of Table 4.2
Subject
HUHGRRY-1
Subject
( DESTRLIHIZE-3 )<\._ HUHGRRY- 1
(CRUSE-5)-- STRUGGLE-1 -( CRUSE-6 )-( STRUGGLE-2)
( HQ-85 ) IDEOLOGICRL
~( HRS-CRRDIHRLITY-2 )-~
( STRUGGLE-1
( CRUSE-6) ( STRUGGLE-2J
( CRUSE-5 ) DESTRLIHIZE-3
(OF-22) PUllER
(STRUGGLE -2 ~( HRS-CRRDIHRLITY-3 )-~
( CRlJSE 6) ( STRUGGLE-1)
A Tourist's Guide to RELATilS 115
knowledge structure, by asking literal and explicit questions of a belief

system or browsing through the structures using the graphical displays of
the belief system examiner. Experiments with analogicalfprecedential retrieval
and inference (Mallery and Hurwitz 1987) are one of several additional modes
of analysis under investigation, of special relevance to Haas's international
conflict management research program.
Reflections on This Case Analysis

Substantive points about this example are also worth making, especially
when it is put in contemporary historical perspective. After Nagy's post-
humous rehabilitation and reburial in 1989, more scholars in Eastern Europe
would accept the United Nations General Assembly's judgment, cited in
Table 4.1, that the popular uprising against Soviet intervention was initially
peaceful, "spontaneous and nationalistic." That even Soviet parliamentarians
have condemned the Brezhnev Doctrine interventions speaks to the inter-
national evolution of a shared "noninterventionist" or "national self-deter-
mination" standard of conduct. But to declare liberal internationalist "prin-
<;iples" the victors of the Cold War would require much richer text analysis
oriented toward eliciting and testing the warranting presuppositions of such
various political arguments.
By encouraging us to explicate the background assumptions of any text
and compare it with other perspectives, RELATUS is designed to recognize
and adjust for such inevitable interpretive "prejudices." Surely there are no
completely theory-neutral data vocabularies. But neither should the important,
intentional, interpretation-embedded, deeper "truths" of the Butterworth-
Scranton account be ignored. Brzezinski's or the General Assembly's inter-
pretations, as interpretive and explanatory hypotheses embedded in an
historical account, would not be considered events by most events data-
making practices; neither do those practices code for the cognitive or rational
content of political speeches and texts. Thus event codings based only on
the events in either the original Butterworth-Scranton account or the "literal
and explicit" version in Table 4.2A would not catch either Brzezinski's
fascinating interpretation nor the historically significant, delegitimating "word
force" of the General Assembly's aforementioned motivational interpretation
of behavior by popular elements in the pre-intervention Hungarian situation.
Additional reflections on these analytical findings are appropriate. First
of all, one should think of the ways in which Hungarian, Czechoslovakian,
and other 1989 Eastern European "revolutions" relate to Gorbachev's own
restructuring initiatives, his perestroika; with appreciation for the creative
modes of change all over Eastern Europe, parallels and differences in "de-
Stalinization processes" need to be further investigated abductively, i.e.,
searching for larger generalizations that would fit both processes. The
analytical trick is to try to infer the same "textual causal model" from the
two process clusters, several decades apart, with vastly different outcomes.
Except for the Romanian revolution and ethnic violence mostly inside the
Soviet Union, both changes have at least begun in largely nonviolent ways;
events in the Soviet Union appear to have initiated both processes. But the
later case appears to have more profound consequences for international
order. Why? What are the reciprocal causal relationships between the Soviet
Union and the smaller Eastern European countries not pictured in Figure
4.3?
Second, the tremendous political significance of the 1989 revolutions is
better understood juxtaposed with knowledge of the General Assembly's
determination, by a largely frustrated pro-Western majority, that the Hungarian
uprising was "spontaneous," "nationalistic," and originally peaceful. One
can find elsewhere similarly contested arguments about Czechoslovakia's
"Prague Spring"; evidently, such language is crucially relevant to political
and scientific judgments of the nature and quality of the Cold War and
post-Cold War order in Eastern Europe.
Finally, whereas the Brzezinski-inspired structural analyses of 1956 events
suggest similar, rapid, highly interconnected, contradictory developments in
subsequent Eastern European crises, we should be sensitized to the extent
to which the success of earlier writings by Brzezinski and others on the
totalitarian character of the Soviet system and bloc structure was one reason
for the uniform failure of Western scholars to predict the events of 1989.
This argument (suggested by Hoffmann 1990) instantiates Havel's warnings
about the dangerous, ambivalent, and sometimes even lethal qualities of
language. Indeed a view of international politics that does not deeply investigate
the moral 1political temper of domestic and transnational "civil society"-
a temper we now see in Eastern Europe to have been full of rage and hope,
despite Western "totalitarian" Iabeling-will not correctly predict future
transformations like those of 1989. If these reflections suggest ways in which
RELATUS-based text modeling and analysis go beyond conventional statistical
practices, they also suggest how far its future developers must go before
they can claim adequately to have reconstructed some of the most interesting
analytical-interpretive procedures of contemporary text and discourse anal-
ysis.
RELATUS's Linguistic/Interpretive Modeling Philosophy
Parsability and Constraint Posting as Gateways to Efficient,

Domain-Independent, Context-Sensitive Text Modeling
Heretofore, most content analyses (Weber 1985) and events data research
have treated codifications derived from text "quantitatively"-as "indepen-
dent" observations to be cross-tabulated or correlated. Because of its

coreferential capabilities, the text-modeling process of RELATUS (and similar
systems exploring syntactic, semantic, or pragmatic textual-interconnections)
is very different; moreover, the results of its modeling processes are
hypothesized to be valid across many different content domains, and even
different human languages.
RELATUS's domain-independent applicability derives from a Chomskyean
view of the relative autonomy of the basic syntactic distinctions used by
its parser. This does not preclude the pragmatic and historically reflective
study of interactive word practices within its grammatical coverage.
Context sensitivity, a fundamental property of human action, is linguistically
encoded by RELATUS users in a number of ways. Both culturally general
ontologies, meanings, beliefs, and known truths, and those specific to
particular actors can be represented by explicit, comparable, and modifiable
background texts. Because RELATUS was designed to investigate and
interrelate the interpretive perspectives and the meaningful interactions of
different actors, and to study the conditions for and possibility of their
differentiation and accommodation, including the emergent constitution of
different meanings, identities, beliefs, and practices, it does not presuppose
uniformity in these respects. Moreover, the effects of different histories on
descriptions, interpretations, policy claims, and identifications can be easily
explored by including relevant additional texts in the construction of more
complex text models. Because they are not reduced positivistically to a
small number of relational primitives-as in Schank-Abelson conceptual
dependency models of belief systems (Aiker 1975; Schank & Abelson 1977)-
associative connections for words may change through time and be recognized
as such.
There are two senses in which this approach is a highly efficient one.
First, computational speed: work with Butterworth texts and SPD protocols
suggests that already debugged, parsable sentences reparse in about 1.6 to
2.2 seconds. The eight dense pages of augmented Hungarian intervention
text, when already "debugged," takes about six minutes to be reparsed and
semantically represented. Based on a 4-megaword (16-megabyte) Symbolics
3640, these numbers multiply up to a pretested text-parsing rate of something
like eighty pages per hour. The newest "Ivory" LISP machines enhance
these numbers by a factor of about five.
But here is where the second kind of context-sensitive efficiency comes
in, extensibility. New concepts and relations can be efficiently connected
into the text models without putting aside existing connections among graph
elements. Purposive, interpretive, figurative, or precedential connections
among events from different partial perspectives could be mapped onto the
same event-based, linguistically meaningful text model; this would be the
case, for example, if the RELATUS model's graphical representation of the
de-Stalinization process of Eastern Europe were somehow mapped into an

historically updated belief system capable of reflectively modifying that
dynamic model in creative, autonomy-enhancing ways. Thus the ceteris
paribus reasoning of contemporary statistical practice can be transcended
by interpretively more explicit studies of meaning construction, history
remembering, strategic action planning or argumentative practice (Mallery
and Hurwitz 1987; Alker 1988).
RELATUS's Evolutionary Constructivist Epistemology

Epistemologically, Mallery and Duffy describe their extensible approach
to semantic representation as "lexical-interpretive semantics" because it uses
only original lexical items (words) and syntactically canonical representations
as a starting point for constructing flexible semantic categorizations (Mallery
and Duffy 1990; Mallery, Hurwitz, and Duffy 1987). Neither logical positivist
nor platonic meaning primitives are thus assumed. The as yet only partially
achieved goal of RELATUS's developers is not to require that meaning
congruences all be determined in a statical, categorical, definitional way.
" 'Lexical-interpretive' semantics differs from approaches relying on [Vale-
style) semantic universals because meaning equivalences are determined
dynamically at reference time for specific-language users with individual
histories, rather than statically in advance for an idealized-language user
with a general but unspecified background knowledge" (Mallery, Hurwitz,
and Duffy 1987).
The RELATUS system is a calculated bet that informal, modifiable
representations and bottom-up generality are productive ways to computerize
social-science-relevant texts. Some time in the future, RELATUS developers
hope that their multiple belief systems can reflectively induce, or hypothetically
abduce, the categories in terms of which individual beliefs are defined and
revised.
Learning from the Past

A long time ago Dilthey conceived of hermeneutics as the interpretation
of texts, and hermeneutic philosophy as the development of objective methods
for confirming human intersubjectivity through a critical interpretation of
human expressions (Aiker, Lehnert, and Schneider 1985: 36). Taking a
different interpretive tack, Weber distinguished social science from physiology
in terms of the meanings that human beings infuse their behavior with; he
argued that interpretations of intentionality must precede explanatory analyses.
Vico, like Aristotle and lbn Khaldun before him, and Marx and Havel more
recently, has stressed the artificial, the contingent, the creative, and sometimes
painful quality of social, economic, and political practices.
An important motivation for the exploration of "computational herme-

neutic" approaches to scientific analysis is the belief in the importance for
cumulative scientific work of the largely qualitative, textually encoded in-
formation, intertwined with interpretively suggestive but debatable (or "sub-
jective") insights, found in so many of history's primary sources, literature's
imaginative fictions, and the social sciences' experimental protocols, open-
ended narratives, scholarly narrative accounts, and its classical works.
We have now briefly illustrated RELATUS's hermeneutically sensitive, post-
positivist, never perfect, but still potentially revolutionary, data-making and
analysis potential. Although statistical analyses of words, texts, and actions
are of course possible and useful in various ways, we believe the extensible,
lexical, relational, graphic representations shown above point toward more
realistic, creativity-respecting foundations for the textually oriented social
sciences of the future. Being excited about this vast, new, perspective-
sensitive, meaning-rich world of replicable and testable inquiry into the
mediating and constitutive functions of language, we are eager to have
others join us in the development of rigorous new methods for productively
analyzing tens and hundreds (or even eventually, in parts, thousands) of
pages of theoretically relevant, empirically critical, and practically important
textual accounts of human conduct. Because the sunk costs of taking words
seriously are even higher than those associated with discipline-rewarded
statistical practices, we are optimistic that a better mix of these skills will
evolve in the future of political science and the other social sciences. 12
Our special concern has been with texts that interpretively report on
conflictful and cooperative political interactions through time, and the enduring
institutions they sometimes create, other times conceal, and even less
frequently transcend. The present example has only begun to suggest how
important words are in the historical deeds and social processes we identify
as the creation, reproduction, and transformation of international political
orders.
A textual revolution in data making has not yet swept the international
politics discipline-and it will not until text-modeling procedures are less
expensive, more compellingly validated, and taught in graduate research
departments. We, at least, find the textual representations of this paper
more aesthetically appealing than the intentionally and linguistically impov-
erished representations of contemporary political statistics; perhaps that is
a telltale omen of a more human, lifelike, and creative future for rigorous
social science.
Notes
This a revised version of Alker, H. R., Jr., and J. C. Mallery, "From Events Data
to Computational Histories: A RELATUS-based Research Program on the Collective
Management of International Conflict," presented at the 1988 Meeting of the

International Studies Association, April 1988.
1. This chapter synthesizes and updates presentations by the authors at the
conference on New Technologies for Events Data Analysis, held at MIT November
13-15, 1987. The contributions to those remarks, and these, of Frank Sherman,
the principal creator of the SHERFACS quantitative events data set used here, are
gratefully acknowledged; the textual account of the 1956 Hungarian crisis in Table
4.1 (Butterworth, with Scranton 1976), reprinted with the permission of Robert Lyle
Butterworth and the University Center for International Studies at the University of
Pittsburgh, has also been vital.
Although the views expressed here should not be identified with those of our
many sources of partial support, we gratefully acknowledge them here. The Data
Development for International Research project of the University of Illinois, with
funds from the National Science Foundation, cosponsored the New Technologies
conference with the Center for International Studies at MIT; the MIT Artificial
Intelligence Laboratory made its facilities available for that purpose and much of
the RELATUS system development and application work reported on here. The
laboratory is partially supported by the Advanced Research Projects Agency of the
U.S. Department of Defense under Office of Naval Research, contract number N00014-
85-K-0123. A grant for research on international security and arms control from
the John D. and Catherine T. MacArthur Foundation helped fund the conference
and RELATUS applications to U.N. conflict management activities and prisoner's
dilemma game protocols. Components of RELATUS were developed by Duffy and
Mallery during summers at Gould, Inc., Semiconductors Division, and at Symbolics,
Inc. Work done by Duffy from September, 1985, until August, 1989, has been done
at the University of Texas at Austin. Symbolics, Inc., has supplied equipment support
for Duffy since December, 1986. Recently, Duffy's work has been at the New York
State Center for Advanced Technology in Computer Application and Software
Engineering, Syracuse University, and Mallery has been helped by NSF award DDM-
8957464 to D. Sriram, Department of Civil Engineering, MIT
2. The best previously available overview descriptions of RELATUS are contained
in Duffy's Ph.D. dissertation (Duffy 1987, especially the appendix by Duffy and
Mallery) and in Mallery's S.M. thesis (Mallery 1988a). Copies are obtainable at cost
from the Institute Archives and Special Collections Office, 14N-118, MIT. The first
publication containing some of the key ideas of RELATUS is Duffy and Mallery
1984. The present paper attempts a more up-to-date, but simplified, "tourist's" guide
to a system that is still evolving. The alphabetical order of authors takes into account
Alker's principal role as drafter of the present text, and Hurwitz's contributions as
a collaborator in the work reported here and as RELATUS's most intensive new user
and internal critic.
3. The basic RELATUS natural language environment is implemented by 3.5
megabytes of COMMON LISP source code, whose separate parts discussed in this
chapter are copyrighted, 1983-1990, by their creators, Gavan Duffy and/or John
C. Mallery. Under various copyright restrictions, additional, loadable, parsing and
analysis systems have been, or are being, developed.
4. From the perspective of the Haas-Butterworth-Aiker-Sherman research program,
the major purpose for creating such formally structured and manipulable narratives
A Tourist's Guide to RELATOS 121
is to make possible their comparative treatment as differently defined, selected, and

synthesized precedents in a process model capable of simulating alternative past
and future pathways of U.N. conflict management practice. Earlier stages in our
textual modeling of U.N. "precedent logics" or "reflective logics" are reviewed in
Alker, Bennett, and Mefford 1980; Mallery and Hurwitz 1987; and Mallery 1988a.
Mefford's chapter on learning models in the present volume may be considered
complementary to these earlier efforts.
The need to synthesize further "narrative" and "quantitative" case accounts has
been apparent in this research program ever since Haas put brief textual summaries
of cases in the center of his McBee sort cards; it is brought more glaringly to the
reading public's attention by a comparative glance at the textual "data summaries"
and brief quantitative codings at the end of Butterworth, with Scranton, 1976. The
question of the degree to which textual connectedness and the contextual inter-
pretability are destroyed by conventional statistical data making has also to be
answered.
An independent account of the ontological commitments of these paradigmatically
different data-making approaches is suggested by Spencer 1982 and Alker 1986.
Behavioristic "statistics," naturalistic "theory," "history," "philosophy," market
"games," and "language practice" (Havel's "words") exist in different but related
domains. We look forward to the day when the difficulties raised by such profound
differences are no longer typically sidestepped or avoided by paradigmatic fiat.
5. Because the Haas research program has its own substantive and practical
research purposes, it need not be thought of as the best demonstration domain for
each of RELATUS's methodological capacities. Although the exceptionally explicit
causal reasoning of the text example in this paper is usefully rich for present
illustrative purposes-which include rethinking the way certain institutional "success"
and "failure" coding practices vis-a-vis collective security-seeking activities have
been previously defined-the parsing examples we shall give are extremely simple
ones. In attempting to integrate the narrative and quantitative data of SHERFACS,
Mallery (1988b) has found serious problems as well in the specificity with which
individual actors within collective agents have been regularly identified.
Nor are Butterworth-Scranton summaries adequately explicit concerning the
theoretical, historical, or practical debates of recent international politics. No one
has yet done the essential work of getting comparable, historically-culturally-politically
distinct accounts of the "same" past crises management situations and resulting
diplomacy so that the multiperspective, reality-constituting "diplomatic word work"
of composing peace-bringing resolutions, getting them passed, and ensuring their
implementation can be more fully comprehended and simulated.
6. "Literal" language refers to text in which all words are used according to their
dictionary definition; also, no figures of speech or tropes like irony and metaphor
are allowed. At present RELATUS allows only one such definition per word; consistent
with this limitation, note the ambiguities concerning "state" in the comments part
of Table 4.26. "Explicit" language, in our usage, contains no implicit premises or
referents that the reader is required to infer. These serious limitations mean that
many teleological, deductive, hypothetical, and figurative forms of deliberative reference,
important parts of human natural language referential competence, are now only
partially implemented.
On the other hand, these limitations made possible the successful text modeling
of the full version of Table 4.2, which is especially sensitive to motivational and
causal "for" or "because" relationships therein, treated in a manner corresponding
to Winston-Katz practice. (See Winston, 1980; Katz and Winston 1982; Katz 1990.)
With this background information, much of what was done in transforming the
Butterworth-Scranton narrative of Table 4.1 into the literal and explicit text in Table
4.2A should make sense. The progressive extension of the implemented language
model would be an important part of a community-shared, RELATUS-based, research
program.
7. Polimetrically, the same point can be made concerning similar natural language
processing systems like those of Winston and Katz; recent Yale text-modeling
literature, selectively reviewed in Riesbeck and Schank 1989; Dyer 1983; Kolodner
1984; and Birnbaum 1986; or the ambitious CYC project (Lenat and Guha 1990).
Along with the technical computational literature, political scientists will need to
become expert in "text linguistics," the formal representation of "commonsense"
knowledge, culturally sensitive "pragmatics," and "discourse analysis." (See de
Beaugrande and Dressier 1981; Levinson 1983; Stubbs 1983; Minsky 1986; Shapiro
1988; Polanyi 1989, for starters.)
8. Besides omitting the Brzezinski interpretation of the significance of the earlier
Nagy period in the first paragraph of the Butterworth-Scranton text, Mallery's "literal
and explicit" reformulation omits the meaning-loaded reference to " 'intellectual'
Petbfi Circles" in the second paragraph. Such omissions can, in principle, be corrected
for by embedding particular textual accounts in larger RELATUS belief network
representations where such terms or references are explicated.
Mallery and Duffy's demonstration of a slightly earlier version of a parsed Hungary
case text at the 1984 AAAI meeting was, we feel, an important success in the
natural language processing and knowledge representation concerns of the Artificial
Intelligence Laboratory. The related two papers for International Studies Association
and Political Science audiences (Mallery 1987; and Mallery and Hurwitz 1987) also
represent, we believe, progressive problem shifts in the "generalized precedent logics"
research program of (Aiker, Bennett, and Mefford 1980).
9. We have in mind the ideal of "full information maximum likelihood statistics,"
as it is used, inter alia, in King 1989. Note that the variety of interconnective
constraints among words, phrases, sentences, and ideas in textual "data networks"
means that the null models of much "correlational" statistical investigation (and the
categorical or linear likelihood distributions too quickly adapted from nonhuman
contexts) look to us and to most linguists like mutilations usually to be avoided.
Because they are not investigating meaning, Mosteller and Wallace's ingenious adaption
of Bayesian statistical methods to discriminate among the unconscious, minor-word-
using propensities of the authors of the Federalist papers does modestly help resolve
authorship controversies. An adequate statistics of meaning and action structures
awaits linguistic progress in representing meanings (Aiker 1975).
10. The full figure, more than twice the size of Table 4.2C, is too large to fit
legibly onto a single page. In it only mammals and organizations are instantiated as
(living) systems, with the former limited to the "leaders" of Table 4.2C, and the
latter differentiated into governments, bureaucracies, parties, countries, and corn-
munism or capitalism as modes of being. The full "literal and explicit" text
corresponding to Table 4.1 takes five rather dense, small-print pages, and the
experimental background file takes three, totalling perhaps ten to twelve printed
pages of text. The belief system representation for these eight dense or ten to twelve
normal printed pages is about 450 kilobytes worth of syntactically, semantically
coreferenced graph. This background-specific representation surely transcends Haas
and Butterworth's earlier single-card-per-case data storage (containing perhaps 120
bytes, or roughly 1/4000th as much information!).
11. What kinds of "literal and explicit" sentence inputs can be successfully
"parsed" and "referentially integrated" by RELATUS depends, of course, on the
current system implementation. A precise metric for describing RELATUS's parser's
scope has not yet been found. In a fast computing environment, Mallery has informally
suggested the writing of a "grammar checker" matching candidate inputs against
a continually updated "present history" of past parsingfcoreferencing successes.
Duffy has made the more radical Piagetian suggestion, also nonimplemented at
present, of incorporating the parser into the GNOSCERE belief system so that
learning about its successes and failures could be a more integral part of RELATUS's
future development. At present, parsing and coreferencing "bug (failure) reports"-
important evidence for the scientific and technical assessment of system perfor-
mance-are easily (some automatically) reported by system users to its developers.
12. Work in computational linguistics or text modeling and analysis presupposes
some knowledge of formal language theory and of grammar-modeling-relevant
computational languages such as PROLOG or LISP. Social science students would
do well to begin with a review of the ways in which Chomskyean formal grammars
subsume Skinnerian Markov models of verbal behavior (Chomsky 1957, 1963); they
will find an easy introduction to the SCHEME dialect of LISP in Eisenberg 1988,
and an authoritative overview of COMMON LISP in Steele 1990. The most ambitious
text modeling of encyclopedic knowledge is the CYC project (Lenat and Guha 1990).
It can be put in contrasting philosophical perspectives by reading Winograd and
Flores 1986 and Minsky 1986.
Bibliography
Abelson, H., Sussman, G., with Sussman, J. Structure and Interpretation of Computer
Programs. Cambridge, Mass.: MIT Press, 1985.
Alker, Jr., H.R. "Polimetrics: Its Descriptive Foundations." In F. Greenstein and N.
Polsby, eds., Handbook of Political Science, Vol. 7, Ch. 4, pp. 139-210. Reading,
Mass.: Addison-Wesley, 1975.
- - - . "Bit Flows, Rewrites and Social Talk." In Todai Symposium 1986, Information
and Its Functions. Tokyo: Tokyo University, Institute of Journalism and Com-
munications, 1986.
- - - . "The Dialectical Logic of Thucydides' Melian Dialogue," American Political
Science Review, Vol. 82, No. 3 (1988), pp. 805-820.
- - - . "Emancipatory Empiricism: Toward the Renewal of Empirical Peace Re-
search." In Peter Wallensteen, ed., Peace Research: Achievements and Challenges,
Boulder: Westview Press, 1988.
- - - . "Fairy Tales, Tragedies and World Histories: Towards Interpretive Story

Grammars as Possibilist World Models," Behaviormetrika, No. 21 (1987), pp. 1-
28.
Alker, H.R., Jr., and Ashley, Richard K. After Neorealism: Anarchic Institutions of
World Politics. In preparation.
Alker, H.R., Jr., Bennett, J., and Mefford, D. "Generalized Precedent Logics for
Resolving Insecurity Dilemmas," International Interactions, Vol. 7 (1980), pp. 165-
206.
Alker, H.R., Jr., and Sherman, F.L. "Collective Security-Seeking Practices Since
1945." In Daniel Frei, ed., Managing International Crises. Beverly Hills: Sage,
1982.
Alker, H.R., Jr., Lehnert, W.G., and Schneider, D.K. "Two Reinterpretations of
Toynbee's Jesus: Explorations in Computational Hermeneutics." In G. Tonfoni,
ed., Artificial Intelligence and Text-Understanding: Plot Units and Summarization
Procedures. Parma: Edizioni Zara, 1985.
Birnbaum, L. Integrated Processing in Planning and Understanding. New Haven: Yale
University Computer Science Dissertation #489, December 1986.
Butterworth, R., with Scranton, M. Managing Interstate Conflict, 1945-1974. Pittsburgh:
University of Pittsburgh, Center for International Studies, 1976.
Campanella, M. Between Cognition and Rationality. Bologna: A. Meunier, 1988.
Carbonell, J. Subjective Understanding: Computer Models of Belief Systems. Ann
Arbor, Mi.: University of Michigan Research Press, 1981.
Chomsky, N. Syntactic Structures. The Hague: Mouton, 1957.
___ . "Three Models for the Description of Language," (and other chapters co-
authored with G. Miller) in R. Luce, R. Bush, and E. Galanter, eds., Handbook
of Mathematical Psychology, Vol. 2. New York: Wiley & Sons, 1963.
de Beaugrande, R., and Dressier, W. Introduction to Text Linguistics. New York:
Longmans, 1981.
Duffy, G. "Categorial Disambiguation," in AAAI-86: Proceedings of the American
Association for Artificial Intelligence, pp. 1057-1061. Philadelphia, August 1986.
___ . Language, Politics and Method: Grounding a Computational Hermeneutic.
Ph.D. dissertation. Cambridge, Mass.: MIT Department of Political Science, 1987.
Duffy, G., and Mallery, J. "Referential Determinism and Computational Efficiency:
Posting Constraints from Deep Structure," in Proceedings of the 1984 National
Conference on Artificial Intelligence, August 1984.
Dyer, M. In Depth Understanding. Cambridge: MIT Press, 1983.
Eisenberg, M. Programming in Scheme. Redwood City, Calif.: Scientific Press, 1988.
Haas, E.B. Collective Security and the Future International System. Monograph Series
in World Affairs, Vol. 5, No. 1 (1967-68). Denver: University of Denver, 1968.
___ . Why We Still Need the United Nations: The Collective Management of
International Conflict, 1945-1984. Policy Paper #26 in International Affairs. Berkeley:
University of California, Institute of International Studies, 1986.
Havel, V. "Words on Words," written acceptance speech for the Peace Prize of the
German Booksellers Association, awarded October 15, 1989. English translation
published in The New York Review of Books, Vol. 36, Nos. 21 and 22, pp. 5-
8.
Hoffmann, S. "A Plan for the New Europe," The New York Review of Books, Vol.
36, Nos. 21 and 22, pp. 18-21.
Katz, B. "Using English for Indexing and Retrieving." In P. Winston, ed., Artificial
Intelligence at MIT: Expanding the Frontiers. Cambridge: MIT Press, 1990.
Katz, B., and Winston, P.H. "A Two-Way Natural Language Interface." In P. Degano
and E. Sandewall, eds., Integrated Interactive Computing Systems. Amsterdam:
North-Holland, 1982.
Kelman, K., ed. International Behavior. New York: Holt, Rinehart & Winston, 1965.
Keohane, Robert 0. "International Institutions: Two Approaches," International Studies
Quarterly, Vol. 32, pp. 379-396.
King, G. Unifying Political Methodology. New York: Cambridge University Press, 1989.
Kolodner, J. Retrieval and Organizational Strategies in Conceptual Memory: A Computer
Model. Hillsdale, N.J.: Lawrence Erlbaum, 1984.
Lenat, D.B., and Guha, R.V. Building Large Knowledge-Based Systems: Representation
and Inference in the Cyc Project. Reading, Mass.: Addison-Wesley, 1990.
Levinson, G. Pragmatics. New York: Cambridge University Press, 1983.
Mallery, J.C. "Universality and Individuality: The Interaction of Noun Phrase Deter-
miners in Copular Classes," in Proceedings of the 1985 Annual Convention of
the Association for Computational Linguistics. July, 1985.
___ . "Computing Strategic Language: Natural Language Models of Belief and
Inference," paper given at International Studies Association Annual Meeting, 1987.
___ . Thinking About Foreign Policy: Finding an Appropriate Role for Artificially
Intelligent Computers. S.M. thesis. Cambridge, Mass.: MIT, Department of Political
Science, 1988.
- - - . "Constraint-Interpreting Reference," AI Memo No. 827. Cambridge, Mass.:
MIT Artificial Intelligence Laboratory. Forthcoming, 1990. Available in draft since
1985.
Mallery, J.C., and Hurwitz, R. "Analogy and Precedent in Strategic Decision-Making:
A Computational Approach," paper given at American Political Science Association
Annual Meeting, 1987.
- - - . "A Computational Model of Semantic Perception," AI Memo No. 799.
Cambridge, Mass.: MIT Artificial Intelligence Laboratory. Forthcoming, 1990. Drafts
date back to 1986.
Mallery, J.C., Hurwitz, R., and Duffy, G. "Hermeneutics." In S.C. Shapiro, The
Encyclopedia of Artificial Intelligence. New York: John Wiley, 1987.
Minsky, M. The Society of Mind. New York: Simon & Schuster, 1986.
Moerman, M. Talking Culture: Ethnography and Conversational Analysis. Philadelphia:
University of Pennsylvania Press, 1988.
Mosteller, F., and Wallace, D.L. Inference and Disputed Authorship: The Federalist.
New York: Springer Verlag, 1984.
Polanyi, L. Telling the American Story. Cambridge, Mass.: MIT Press, Bradford Books,
1989.
Riesbeck, C.K., and Schank, R.C., Inside Case-Based Reasoning. Hillsdale, NJ: Lawrence
Erlbaum, 1989.
Schank, R.C., and Abelson, R. Scripts, Plans, Goals and Understanding: An Inquiry
into Human Knowledge Structures. Hillsdale, N.J.: Lawrence Erlbaum, 1977.
Shapiro, M. The Politics of Representation. Madison: University of Wisconsin Press,

1988.
Sherman, F.L. Part-Way to Peace: The United Nations and the Road to Nowhere?
Ph.D. dissertation. College Park, Pa.: Pennsylvania State University, Department
of Political Science, May 1987.
___ . "SHERFACS: A New Cross-Paradigm, International Conflict Dataset," paper
presented at the International Studies Association Annual Meeting, St. Louis,
1988.
Spencer, M.E., "The Ontologies of Social Science," Philosophy of Social Science,
Vol. 12 (1982), pp. 121-141.
Steele, G.L., Jr. Common LISP: The Language. Norwood, Mass.: Digital Press, 1990.
Stubbs, M. Discourse Analysis. Chicago: University of Chicago Press, 1983.
Weber, R. Basic Content Analysis. University Paper #49. Beverly Hills: Sage, 1985.
Winograd, T., and Flores, F. Understanding Cognition and Computers. Norwood, N.J.:
Ablex, 1986.
Winston, P.H. "Learning and Reasoning by Analogy," Communications of the ACM,
Vol. 23, No. 12 (1980), pp. 689-703.
___ . "Learning New Principles from Precedents and Exercises," Artificial Intel-
ligence, Vol. 19 (1981), pp. 321-350.
5
Reasoning and Intelligibility
James P. Bennett and Stuart J. Thorson
I've always been struck by a story told by Paul Warnke, the former director of
the Arms Control and Disarmament Agency. He said he once spent an hour
briefing Edmund Muskie, then a U.S. Senator from Maine, about the strategies,
counter-strategies, and options of nuclear policy and nuclear war. When he was
finished, Mr. Muskie said to him: "You've got to be nuts."
-Richard Reeves, Sunday Times, January 24, 1986
Politics can be seen as a creative process in which events occur that

are in most every sense novel. From this perspective, a remarkable feature
of political life is that these events, though novel, are often intelligible to
political practitioners. By "intelligible" here is meant nothing more than that
political practitioners can "make sense of" the events. This is interesting
precisely because making sense of novel events requires that political actors
(individuals, bureaucracies, and so on) extend an existing largely intelligible
world to include this new happening. That which we as political analysts
term a "decision" is parasitic upon the particular discretization of phenomena
provided by the intelligibility extension. How this extension is carried out
is important to our understanding of politics because this extension is in
most all cases underdetermined. Which extension may well determine
important political possibilities.
A brief example may help to clarify. When President Kennedy was first
shown a U-2 photograph of Cuba he is said to have remarked that it looked
like an aerial photograph of a football field. Such an interpretation is one
possible way of extending the existing world to include the photograph.
However, as we know, the intelligence community (in particular the Pho-
tographic Interpretation Center within the CIA) suggested another extension-
namely that the Soviet Union was developing an offensive missile capability
in Cuba. This latter extension was the one adopted and the photograph was
rendered intelligible. This categorization of the photographic evidence carried
with it a whole set of rules that in part served to constitute the event in
127
128 James P. Bennett and Stuart J. Thorson
an intelligible way and were also used to structure the policy debate in what
became understood as the Cuban missile crisis.
Computational theories of politics typically begin by taking as a given
the intelligibility of politics and then attempt to show how political choices,
decisions, or events can be generated. Such theories are, we argue, hard
pressed to account for the "creative" aspects of political intelligibility, as
they all depend upon implementing categorization as a pattern match against
a fixed set of templates. Given a successful match, the specific situation
inherits non-overridden defaults from the template. However, categorization
is not like simple pattern matching, in that it leaves open the possibility
that one side effect of a particular categorization will be to modify existing
categorization rules. In other words, the notion of politics as a creative
activity requires that the way in which categorization is done at some time
may in part depend upon past categorizations.
Note that from this perspective, the notion of an event in international
affairs (or, for that matter, international affairs itself!) is not that of a physical
event that comes along and requires only to be recognized. Rather, it suggests
that what we analysts term an "event" is the result of a complicated and
poorly understood set of intelligibility extensions carried out in social settings
and heavily mediated by social institutions (Aiker 1988). Parenthetically, it
also suggests that forecasting systems that do not have embedded within
them a plausible theory of precisely how these intelligibility extensions are
done are themselves likely to become a component of this extending process
and thus become "event creators" by freezing the extensions of the past
and applying them in a changed future.
Thus at the abstract level our concern is with (1) understanding the
procedures through which things are seen as intelligibly political and (2)
examining the political consequences of these procedures-among which
may well be changes in how (1) is effected. This in turn requires that
attention be paid to questions of how understandings of politics can be
represented and modified. We will illustrate how one might approach these
issues from a computational perspective through an extended discussion of
deterrence.
Deterrence
Computational modeling of nuclear "deterrence" shows the complexity
of the principal concepts underlying deterrence as well as the degree to
which any calculus of deterrence must reflect upon and embody a diverse
set of problematic beliefs about the contemporary political environment.
Chief among these beliefs is the model of beliefs of other parties, such as
those against whom influence is exerted. The discussion is motivated by
the need to illustrate factors affecting the intelligibility of such strategies,
Reasoning and Intelligibility 129
not to suggest that human reasoning somehow fails to master such intricacies.
In fact, every one of the features identified below in our inquiry has been
identified previously in the academic literature about deterrence. What has
not been clearly identified is the relationship of each feature to the whole
concept, or the complexity of the whole. Indeed, the complexity is such
that when heads of state or other politically authoritative figures refer to
"deterrence" in private discussions or public pronouncements, they cannot
intend to mean all that is embodied in the concept as we unfold it. For
instance, when President Reagan spoke of "deterrence without fear" in his
Star Wars speech in 1983, did he have in mind a novel concept of deterrence
without fear for both the United States and the USSR (and, if so, what must
he have understood to be the role of threats in deterrence)? Or did he mean
deterrence without fear for the single party possessing strategic defense? 1
The Internal Structure of "Deterrence"

In the United States, Britain, and France there has been little disagreement
about the core meaning of "nuclear deterrence." Simply formulated, "to
deter" means "to dissuade by threat of reprisal," and "deterrence" is thus
the practice of attempting to dissuade by threat of reprisal. The first clear
statement of deterrence in the nuclear era (Brodie 1946) employed the basic
terms with meanings taken from ordinary language, a practice we continue
here.
"To dissuade" means "to persuade one not to do something" (Wolfers
1962: eh. 7; note also Schelling's distinction between "compellence" and
"deterrence"). The school of speech act analysis (Searle 1969) of linguistic
pragm~ics categorizes "threaten," like "promise," as a performative verb.
This means that one achieves things-one acts-by uttering a formula that
is recognizable within a linguistic community. The features of convention
and community that underlie threats contribute to their problematic efficacy
in international politics (Russett 1963; Jervis 1979).
"Reprisal" denotes an intentionally harmful act. The prefix "re-" limits
the term to an act taken in response to another. Exploring the core meaning
of "deterrence" requires us to examine in greater detail the formula to deter
= to dissuade by threat of reprisal. This formula lacks specification of
agents and objects. Specifying these correctly and consistently throughout
an analysis is trickier than many theorists have realized.
A widespread understanding of the historical practice of nuclear deterrence
holds that one cannot have a policy of deterrence secretly. In fact, one
must utter a threat in order to execute a deterrent act. The party intended
to be influenced must hear and comprehend the threat in order that it
succeed. This understanding makes the linguistic component of deterrence
essential. Extralinguistic efforts to dissuade another party should be labeled
an act of "menacing" (e.g., in international politics by the movement of

troops, or in ontological disputes among philosophers by the wielding of a
fireplace poker). Linguistic encoding is necessary to express a contingent
relation between action by the party-to-be-deterred and sanction by the
would-be-deterrer. Communication of a contingency is required in "deter-
rence"; this is one feature that distinguishes it from "defense." By the
"internal structure" of deterrence we mean the relations constituting "dis-
suade," "threaten," and "reprisal," and the semantic composition of these
three terms to define "deterrence."
Who- Whom Issues: Four Simultaneously Valid Perspectives

Any situation of putative relevance to deterrence can be viewed from (at
least) four perspectives. To simplify exposition, we shall assume that there
exist several politically and cognitively autonomous parties, one of which
attempts to make a deterrent threat. One perspective is that of the party
attempting to deter-let us call this party the would-be deterrer (so that
the success or failure of its action is not presupposed). A second perspective
belongs to the party that the would-be deterrer is attempting to influence.
Let us call the second party the party-to-be-deterred, but note that this
designation is meaningful only from the perspective of the first party. The
second party might perhaps appreciate that the first is attempting to deter
it from some act, in which case it can accept the label of party-to-be-
deterred. A third perspective belongs to the second party when it does not
recognize itself to be the target of a deterrent threat. Finally, a fourth
perspective is occupied by contemporary or later third parties (who may
be "historians"). These parties may of course share the perceptions of either
of the other parties, or may conceive of the situation entirely differently.
Both historical analyses seeking to identify determinants of deterrence
success or failure, and policy prescriptions addressing the management of
deterrent threats should select one of these four perspectives and retain it
consistently. Studies that shift among these or confuse historians' perspectives
with that of the would-be deterrer are bound to produce confusion (see
Levy 1989 for a brief review of empirical studies). "Deterrence" is not a
necessary category of human action. People constitute action as "deterrence"
using their own idiosyncratic perceptual and cognitive capabilities, linking
together phenomena and presuppositions about a situation they can only
incompletely observe. The complexity and relativity of "deterrence" is due
to its continual construction in the minds of parties engaged in hostile
international relations. Until explicitly noted below, this study takes the
perspective of the would-be deterrer. We ask what it must know and do in
order to undertake a deterrent act.
Deterrent Acts Versus Deterrent Policies

Deterrent policies are more complex than deterrent acts. lt is hard to
imagine a sense in which one could be said to have a policy of deterrence
without ever pronouncing (i.e., uttering) a threat. But deterrent acts appear
frequently in historical contexts in which the imputation of deterrent policy
is implausible (as in most all the cases studied by Huth and Russett 1984,
1988). By policy one must mean a generalized intention to pursue some
goal or objective. "Policy" is typically less specific than "plan" -at least
as the terms are used in foreign affairs. "Policy" refers to organized intentions
that persist over time; thus a durative character of policies tends to distinguish
them from acts.
The time span over which polices must be expressed imposes demanding
modeling requirements for linguistic tense and aspect. In particular, it requires
us to construct a computational language for process, in other words, a
language of dynamics. By limiting treatment in this paper to deterrent acts,
we can get by with a language of comparative statics. Nevertheless, this
will prove adequate to demonstrate the constitutive complexity of "deter-
rence."
A Minimal Procedure to Recognize Deterrent Acts. Attempting to write
a computer program for the recognition of deterrent acts-from the per-
spective of the would-be deterrer-helps us to identify those features that
make the intelligibility of deterrence a problem. We begin in the spirit of
"heuristic programming," with little attention to unessential details and none
at all to matters of computational efficiency.
A party wishing to deter another might reasonably be interested in
producing deterrent acts, not merely recognizing them. But we shall suppose
that the two are bound up closely in the same relation as speech compre-
hension and speech production. A party contemplating making a deterrent
threat might well reflect whether its target can understand the contemplated
action as intended. A would-be deterrer might ask two questions about
deterrence: (1) What are the essential features that, by convention, constitute
a deterrent act? (2) What additional features tend to make deterrent acts
effective?
Let us call the first set of features the "technical conditions" (noting
that Searle reserves this term for yet more technical details that we shall
ignore altogether-such as specifying the existence of parties, their em-
ployment of a translatable language, and the like). The second set of features
will be termed "efficacy conditions." lt is convenient to treat both sets in
parallel, because efficacy conditions can often be stated as more rigorous
requirements on technical conditions. Since we adopt the perspective of a
would-be deterrer, the efficacy conditions are those which contribute to this
party's sense of efficacy of deterrent acts. Other conditions, such as those
a historian concludes contribute to the effectiveness of deterrent threats in

general, are not identified in this analysis. These are surely important, but
they are distinct from those pertinent to calculations by the would-be deterrer,
and they are derived in part from information that is not be available to
that party.
Let a denote the would-be deterrer and b the party-to-be-deterred. The
simplest case distinguishes just two parties; we shall add others below. To
tailor threats appropriately, a would-be deterrer needs to know the identities
of the parties it wishes to deter. A third piece of information is an utterance
by a: typically a number of texts including policy papers, speeches, and
other apparently authoritative statements. Let Ua be the text, or "utterance
by a." How a constructed the "raw material" Ua need not concern us here
except to note that this, too, is a constitutive process that most likely
responds in some fashion to a's experience with recognizing (or not) deterrent
acts. S represents the threat of reprisal a communicates via Ua to b.
A first cut at such a predicate (using scheme notation) can be sketched
as:
(define deter?
(lambda (a b Ua S)
(dissuade? a b Ua
(threaten? a b Ua (reprisal? S)))))
For bound variables a, b, S, and Ua, this predicate answers "yes" if it
determines that Ua contains a's well-formed threat of reprisal S against b,
and if that threat is instrumental to a's dissuasion of b. This predicate tells
us little of substance, for all the reasoning done by a is described in
subordinate predicates that remain to be specified.
Belief Worlds as LISP Environments. When we write programs to investigate
the structure of thinking about performing acts such as deterrent threats,
we attempt to make the specifications generally applicable, so that the
processes described apply to many parties over significant historical periods.
Yet we r'eeognize that the thinking of the parties cannot be identical, if for
no other reason than that they differ in knowledge and beliefs. It is necessary,
therefore, to try to make the programs perform appropriately in different
contexts. Lumping together all the information we think a party has available,
let us call this information base a belief system. To determine how (our
model of) any party constructs an answer to the question (predicate) deter?
we must evaluate that expression within a context restricted to the information
available to the party. LISP offers the programmer a convenient construct
for encapsulating information in this way; it is called an environment. We
can think of a party's total set of beliefs as constituting an environment
within which there may be embedded environments representing, for instance,
the party's model of other parties' beliefs. Notationally, we can let Wa denote
a's set of beliefs (its "world"); Wab denote a's model of b's beliefs; Waba
denote a's model of b's beliefs about a's beliefs; and so on. (On some
modeling requirements for reflective thinking, see Alker et al. 1980 and
Alker 1984.)
In computational terms, an environment can be implemented as a list
of pairs: variables together with the values to which they are currently
bound. For instance, the concept of "hostile superpower" might be represented
by the variable hostile-superpower. In an environment attributed to "the
United States," hostile-superpower might evaluate to "USSR." In a different
environment attributed to "the USSR," it might evaluate to "United States."
Every LISP expression is evaluated in an environment:
(eval expression environment)
The sketchy definition of deter? above does not specify environments

for evaluation of its expressions. This should be remedied by indicating that
certain expressions must be evaluated with respect to the beliefs of party
a. Rewriting appropriately yields
(define deter?
(lambda (a Wa b Ua S)
(dissuade? a Wa b Ua
(threaten? a Wa b Ua (reprisal? S)))))
When the relevant environment for evaluation is obvious, the notation will
be omitted. Otherwise, when a procedure describes action by an agent, the
second argument to the procedure identifies the environment in which
evaluation occurs. 2
Protection of Third Parties- ''Extended Deterrence''

Much of the empirical work on the efficacy of deterrent acts has addressed
the protection of third parties by threat of retaliation against an aggressor
(Levy 1989; Huth 1988). "Extended deterrence" in the form of attempts by
the United States to protect Western Europe from Soviet invasion was the
original requirement for nuclear deterrence, and this objective prominently
shaped postwar forces and plans (Friedberg 1984). When the contingency
that party a hopes to avert involves harm by b specifically upon a third
party c, then the content of the threat differs from the case of "direct
deterrence," in which a attempts to dissuade b from doing harm to a. The
formula above can then be expanded to identify roles for the three parties:
a deters b from doing harm to c by threatening b with reprisal S.
Here "b doing harm to c" represents the contingency C. One may likewise
expand the sanction S to identify a fourth role for party d against which
the sanction is to be directed: a deters b from doing harm to c by threatening

d with harm.
This form is relevant when a believes b to value the welfare of d. In
strategic doctrines that distinguish between an autonomous elite leadership
and the mass of a country, b can represent that leadership and d represent
the mass. In part because strategists in the United States worried that Soviet
leadership might expect to survive a nuclear exchange and thus not be
decisively restrained by the prospect of harm to the Soviet populace, the
United States enunciated as doctrine targeting specifically the leadership of
the Soviet Onion during the Carter administration (Rosenberg 1987).3
The important point is that, from a's point of view, these distinctions
are produced by a's beliefs about the actual relations among the parties at
a particular time. Any informed and cognitively sophisticated party a will
tailor the composition of its threat in response to changes in its beliefs
about these relationships. (On historical deterrent doctrines of Western
powers, see Freedman 1981). Consequently, the particular form of deterrent
threat judged communicable and effective at any time derives from the
parties' beliefs about the nature of contemporary relations among themselves.
And the efficacy of deterrent threats depends in significant measure upon
the parties sharing certain of these beliefs, as we shall show below.
Dissuasion
"Dissuasion" involves a persuading b not to do something. Let us denote
that collection of acts by C to stand for "contingency." We cannot speak
properly of a dissuading b unless two conditions are met: (1) The contingency
C has not yet happened, represented
(not (occur Wa C))4
and (2) party a strongly dislikes the future occurrence of C, represented:
(prefer a Wa (alternatives C) C).
Party a might conceivably make a deterrent threat ostensibly to avert the
occurrence of some C but actually to serve as pretext for its own anticipated
hostilities against another party b. In this case the deterrent threat itself-
including the apparent attempt at dissuasion-might be plausible, yet the
attempt to avert C be insincere. In international affairs, parties sometimes
make threats to dissuade others from action just when they believe that
the others do not intend to undertake that action; thus they succeed, and
hope to gain a reputation for effective dissuasion. (See Bundy 1988: eh. 9;
Garthoff 1987, 1989, regarding claims about empty attempts at dissuasion
during the Cuban Missile Crisis.) Consequently, condition (i) is usual but
not essential-it resembles some "sincerity conditions" identified below.
(i) Party a believes that b intends to cause C to happen.

The efficacy of dissuasion by a may be enhanced when b believes that
condition (i) holds. This can be stated, reflectively, from a's perspective in
condition (3e) where "e" stands for "efficacy," as:
(3e) Party a believes that party b believes that party a
believes that b intends to cause C to happen.
This exemplifies how a party's (here: b's) apparent cognizance of a situation
(here: a's belief about b's intent) can form a basis for effective influence
against the cognizant party.
In similar fashion party a's attempt at dissuasion, if sincere, must be
directed at b because a believes b to control the occurrence of contingency
C. A first approximation at representing a's belief that b "is responsible
for" C is (4e):
(4e) Party a believes that b is responsible for C if a believes b plans
for or intends C to happen, or a believes b will be able to avert the
occurrence of C should others cause C to happen.
Since intentions can rarely be estimated with precision or with confidence,
we will not pursue this efficacy condition further here. As above, the basis
for a's dissuasion is stronger if a believes that b believes condition (4e).
Putting the four conditions together we can sketch a predicate dissuade?,
postponing specification of the means of exercising influence:
(define dissuade?
(lambda (a Wa b Ua means)
(and
(not (occur Wa C))
(prefer a Wa (alternatives C) C)
(believe a Wa
(believe b Wab
(believe a Waba
(intend b Wabab
(cause-1 b
(occur
Wababa
C))))))
(believe a Wa
(responsible b Wab C)))))
If we do not know how successively reflective models are constructed by
a, then we must approximate Wabab by Wab. This reduces condition (3e)
to condition (i), for "deeper" processing produces no new information. As
a practical matter, except in certain highly structured and recurrent situations ·

(such as the game of chess), it appears unrealistic to expect gains from
processing further than 2-ply, viz., Waba.
Note, too, that since b's beliefs are opaque to a, a can never access Wb,
Wba, . . . but instead must draw upon Wab, Waba, . . . respectively.
Formulating models of one's opponent is central to strategic thought, though
beyond the scope of this chapter. A historian focusing upon a brief period
can sometimes assume that the parties' pertinent beliefs are well formed
and fixed. Then the predicate believe could be implemented as a simple
procedure to look up data in a modeled belief system organized like a formal
data base (Bennett 1984). Otherwise, in addition to environment-specific
evaluation, one requires a device for quoting the object of belief (see Haas
1986).
Predicate dissuade? lacks its most important ingredient, reference to a
means of persuasion. In the case of "deterrence" that is a threat.
"Threaten" as a Speech Act

Deterrence involves reprisal or retaliation; deterrent threats must com-
municate in some measure the nature of the response and its gravity.
Although the boundary between "warnings" and "threats" must remain
fuzzy in practice, we expect in the nuclear era for parties to describe
sanctions in at least general terms, such as "the gravest consequences."
But heads of nuclear states or their deputies have used regular occasions,
often ritualistic (such as speeches before prodefense constituencies), to
reassert the character of sanctions.
More detailed examination of the predicate reprisal? requires a description
of the previous interactions among the parties. Instead, we explicate the
linguistic act of threatening, which requires the would-be deterrer to supply
much of the same information.
The structure of predicate threaten? can be fteshed out by following the
spirit of Searle's decomposition of the speech act "promise" (1969), which
is in several respects similar. We do not require that there exist a uniquely
correct treatment of these sociolinguistic conventions; we shall limit our
decomposition to essential features, whose absence an ordinary English-
speaking strategic theorist or policy-maker would note as a serious deficiency
of expression. Whether this (or any) term in English possesses an exact
synonym (viz., semantic and pragmatic equivalent) in other languages is
doubtful, but the important issue is whether the range of linguistic action
categorized in English by "threaten" can be accurately described in other
relevant languages. Additionally, we list selected conditions that can contribute
to the efficacy of threats but are not essential to their conventional expression.
Retaining the perspective of would-be deterrer a, we define "threaten"

as the argument:
(define threaten?
(lambda (a Wa b Ua S)
(and
"condition-}"
"condition-n")))
Identifying components of the speech act "threaten" enables us to fill in
specifications for threaten?
The threat must be communicated by a to b (~rhaps through an
intermediary) in a form (viz., language) comprehensible to b. This does not
mean that the threat is understood by b exactly as a intended. We will not
treat here the problem of translation. Further, we must assume that b is
somehow attentive to a's expression, and that b can "remember" its meaning
over a pertinent period. (Both of these conditions are problematic in cognitively
nonunitary actors such as formal organizations and government bureaucracies,
which both a and b are typically presumed to represent (.Morgan 1977).)
"Party a states Ua to b (thus enabling b to extract pertinent information
from Qa and add it to its set of beliefs Wb)." Combining these operations
we write (5)
(state a b Ua)
to mean that a states Ua to b, and (6)
(believe a Wa (comprehend b Wab Ua))
to mean that a believes that b understands Qa. From (6) a can infer that
b introduces the information of Qa into its set of beliefs Wb; consequently,
a can update its model of b's beliefs, denoted Wab, with that information.
In other words, a infers that b received a's communication as intended.
Because fulfillment of the process (by which a updates its beliefs about
information available to b) cannot be assumed, parties making deterrent
threats often express them repetitively through redundant channels.
Again the efficacy of a threat may be enhanced if a can determine that
the threat was accurately comprehended by b and if b realizes that a knows
this. From a's perspective, this is expressed as (7e):
(believe a Wa
(believe b Wab
(believe a Waba (comprehend b Wabab Ua)))).
As a general matter, confidence by a that b understands the situation just
as a does, contributes to a's confidence in the efficacy of its threat. Such
confidence may, of course, be misplaced. However, we will omit detailing

all of the efficacy conditions of this reflective kind.
Contingency and Sanction

Threats relate contingency C to sanction S. During recent U.S. admin-
istrations the relations between contingency and sanction pertinent to strategic
nuclear policy have been reflected by the policy process and its paper
products. This was exemplified during President Reagan's second term. If
we think of the contingency as setting forth a complicated but abstract set
of conditions, including alternative directions, levels, and schedules of attacks
on the interests of the United States and its allies, then the contingency C
was described approximately in a National Security Decision Directive,
NSDD 13, and its clarifying documents. The nuclear sanctions to be taken
in response were set out, again in general terms, in a statement of Nuclear
Weapons Employment Policy, NUWEP82. The relationship between NSDD13
and NUWEP82 is roughly akin to a "map" linking contingency C and
sanction S. The map intentionally did not formulate a one-to-one relationship
between harmful acts by foreign powers and U.S. responses; alternative
preplanned responses were to be presented as options for the commander
in chief, so that some flexibility could be imparted to nuclear operations.
Flexibility, expressed as absence of prior commitment to specific sanctions,
might be thought to lend a degree of risk-enhancing uncertainty to potential
attackers.
The sanction is specified in detail in another related document, the
integrated operation plans, SIOP6. This enumerates destructive retaliatory
(and, perhaps, preemptive) acts. Letting X denote the set of destructive
acts, we can represent the relationship between contingency and sanction
of that period as
C--.. S----+- X,
expressed in
NSDD ----+- NUWEP ----+- SlOP.
Unless these documents are declassified, one cannot assess the degree of
consistency among them. But their structure acknowledges the need in
theory to relate carefully contingencies to sanctions. So far as we are aware,
documents such as NSDD13 have never been conveyed to the Soviets or
other states, so they cannot comprise Ua. Some of the content is selectively
expressed in public as U.S. "declaratory doctrine." These secondhand
accounts of policy contribute to Ua.
Expression of deterrent threats requires, then, that some information that
circumscribes contingencies is included in Ua, and that the party-to-be
deterred believes that the would-be deterrer has prepared sanctions with
which to respond should the contingency (8) be fulfilled:
(believe a Wa (infer b Wab Ua C'))
where C' will typically be similar to but not identical with (secret) C and
(9):
(believe a Wa (believe b Wab (exists (map Waba C'S'))))
such that S' is not too different from S.

Condition (8) says that a thinks that b can draw the intended meaning
from a's statements, with the result of identifying contingencies more or
less as a intends. Condition (9) says that a thinks that b thinks that a has
thought through the issue of potential responses to those contingencies. 5
Properly communicating the sanction S and its specific retaliatory acts X
requires careful construction by the would-be deterrer. The sanction S is
related to discrete acts X by the relation that, if S is implemented in an
environment, then X is produced: (do a Wa S) - X. First, the sanction
must refer to future acts by a (10):
(not (occur Wab X)) ;that is, acts X are already not
;recorded in (a's model of) b's beliefs.
Second, a must calibrate its sanction S, and from this estimate b's image
of the sanctionS', to yield a result that a believes b finds strongly unattractive.
Skipping details of how a comes to believe that b identifies acts X' (similar
to X), we write (11):
(believe a Wa (prefer b Wab (alternatives X') X))
Although we should find a more expressive way to indicate that nuclear

reprisals are more significant than "inferior options," condition (11) is actually
quite strong. Evaluating (alternatives X') yields all acts except those in X'
that are known to b; the expression (prefer b ... ) yields "true" only if all
the acts of (alternatives X') are preferred by b to any act in X'. In many
contexts, this condition is too strong: A party may prefer damage by some
"small" nuclear reprisals to other fates such as governmental collapse.
Condition (11) suffices for a to infer that b will attempt to avoid acts of
retaliation X'. But the potential value of a's information about X', to manipulate
the behavior of b is lost unless the following is added: "a believes that b
believes that, if C' does not occur, then a will not want to implement the
sanction S' and thus will not do acts X'." (The prime notation is required
because a must anticipate b's estimate of a's intent and capability.) Thus
condition (12) is:
(believe a Wa (believe b Wab

(if (occur Waba C')6
(prefer a Waba X' (alternatives X'))
(prefer a Waba (alternatives X') X')))))
Note that, if C occurs, Waba in (12) will differ significantly from Wa in (11):
Some second-level predicates are designed to create computational "side
effects" that change the situations as a deliberates about effects of hypothetical
future acts. Executing the sanction must be costly to both parties: to the
party-to-be-deterred so that party avoids triggering the contingency; to the
would-be deterrer so that it does not apply the sanction "in the normal
course of affairs." Conditions (11) and (12) express this from a's point of
view (only).
Some analysts of the speech act "threaten" add a "sincerity condition,"
which states that, "a does not intend to apply sanction S unless the
contingency C occurs." lt would seem linguistically proper, however, to
speak of a party making deterrent threats insincerely, with every intention
of striking its opponent anyway. The preferences of party a might change,
so that assessment of its sincerity sometimes reduces to assessment of its
intentions at particular times. These judgments need not be introduced into
the definition of "deterrence," although the would-be deterrer should wish
to appear sincere (regardless of real intentions) and should think itself more
likely effective in threatening if it believes the party-to-be-deterred thinks
that it is sincere. The efficacy conditions can be stated: "Party a believes
that party b believes that, unless C, a does not intend to apply sanction
S." So condition (13e) is:
(believe a Wa (believe b Wab

(if (occur Wab C) oo future
(intend a Waba
(do a Waba S))
;else
(not (intend a Waba
(do a Waba S))))))
This formulation is deficient because it does not explicitly indicate that

(intend a . . .) is to apply over an indefinite future. We have not introduced
tense markers to simplify the exposition. The formulation is also stronger
than the quotation above: lt states that, if C occurs, then and only then
does a intend to do S. lt is thus partially redundant with condition (12). 7
The corresponding "sincerity condition" is simply, "Unless C, a does not
intend to apply the sanction."
As a reasons, satisfaction of condition (12) is justification for party b to
expect that it is possible to avoid the sanction. Condition (13) expresses
an essential feature of threatening: preference reversal caused by the oc-

currence of the contingency. The art of creating commitments is devoted
largely to inducing b to accept a's claims about reversal of its preferences
for implementing the sanction. One path to establishing commitment is to
increase the costs of other options after the occurrence of C. Another is
to increase the automaticity of implementing the sanction if C occurs.
Institutionalizing a deterrent posture by constructing military forces and
communications to facilitate reprisals, and practicing (or, in psychological
terms, "desensitizing oneself to") the use of systems for reprisals, tends to
increase the credibility of commitment, especially when one's opponent can
witness the preparations. Attainable commitments fall short of denying all
choice to a, as in the limiting case of Herman Kahn's hypothetical doomsday
weapon.
Yet the power of commitment by a depends upon b following one line
of reasoning to the conclusion expressed in condition (13). This may be
paraphrased as, "If C occurs, then a intends to implement the sanction
because only in that case are all alternatives to implementing the sanction
less desirable." It seems to us that, for all but the simplest deterrent threats
made in clear-cut situations familiar to both parties, imputing to b the
willingness and ability to reason thus can never be confidently justified. Put
differently, situations in which deterrent threats would seem to be reasonable
acts are unlikely to be situations in which fulfillment of this condition can
be taken for granted. In exception, we note that parties can become
comfortable and confident with policies of deterrence; the institutionalization
of deterrent postures obviates the requirement to utter deterrent threats
very often. Then the weakness emerges that "frozen" threats, concretized
in standard procedures of political-military operations, fail to adjust to
changes in the world that undermine other conditions for successful deter-
rence.
Another efficacy condition relates to a's understanding that, by uttering
the threat, it assumes an obligation to implement the sanction should the
contingency arise. This relates as well to calculations about one's reputation
should one be found to be bluffing. Condition (14) states:
(believe a Wa (cause-2
(state a b Oa)
(believe b Wab
("condition (13)"))))
where the quotation marks indicate that condition (13) should be inserted
with appropriate adjustments to the indices of environments.
From this the would-be deterrer can infer that the party-to-be-deterred-
as well as other parties attentive to the threat-recognizes an obligation
accepted by the would-be deterrer whether or not it is sincere. The obligation
is twofold, as expressed by condition (13). Though it relates intimately to

the historical efficacy of deterrent threats, we shall not pursue further here
either bases for reputation or "audience effects" of deterrence.
The conditions enumerated above (plus some theoretically unimportant
but computationally essential conditions) can be collected to write a program
to identify deterrent acts. Our top-level decomposition of the internal structure
of the concept "deterrence" leaves unspecified procedures such as infer,
cause-1, and alternatives. A second-level decomposition deals with the logical
structure of practical reasoning; a third-level decomposition addresses sub-
junctive reasoning and what it means to hold beliefs about hypothetical
states of affairs, which, if current practice is successful, will never be
realized. These are important philosophical and methodological issues, but
even the top-level expansion reveals enormous complexity to the concept
of "deterrence." If "deterrence" may be unintelligible in most political
discourse and only approximated in practice, then some key applications
of deterrent strategies are more doubtful yet.
Mutual Deterrence
Mutual deterrence is the state of affairs existing when each of two parties
dissuades the other from certain acts by threatening reprisals against the
other. Just in case
(and (deter a Wa b Ua Sa) (deter b Wb a Ub Sb))
an observer distinct from a and from b can conclude that mutual deterrence
is realized. This is straightforward to write but much harder to appreciate.
In fact, it is not a valid expression from the perspective of party a, for Wb
is not observable. That perspective requires a further examination.
We can assume that, if
(deter a Wa b Ua Sa)
then a can infer
(believe a Wa (deter a Wa b Ua Sa))
because for such self-conscious activity as uttering felicitous deterrent threats

one typically understands what one is trying to do. From a's perspective
"a believes, by having properly uttered a deterrent threat, that a deters b."
Further, if a perceives that b properly uttered a deterrent threat that a
acknowledges actually dissuaded it, a can conclude "a believes, because b
properly uttered a deterrent threat, that b deters a." From these two facts,
one can conclude that "a believes that mutual deterrence holds between a
and b."
But that is a very different thing from saying (as in the perspective of
a third party) that in fact mutual deterrence describes the state of affairs
relating a and b as they understand it. For this conclusion, one requires
the two parallel conditions taken from b's perspective. In other words,
"mutual deterrence" is realized when four conditions are simultaneously
satisfied (which we write out to emphasize the complexity lurking beneath
this simple label) Thus, condition (15):
(and
(believe a Wa (deter a Wa b Ua Sa))
(believe a Wa (deter b Wab a Ub Sb))
(believe b Wb (deter b Wb a Ub Sb))
(believe b Wb (deter a Wba b Ua Sa)))8
It is tempting to challenge claims about "the" contemporary nuclear
relationship between the United States and the USSR. If one asserts that it
is "mutual assured destruction" (as a third-party's description; certainly not
as official policy), then one must be prepared to demonstrate that all of
condition (15) is satisfied. Even if true, it is very hard to demonstrate this
at politically acceptable levels of confidence. But most readers, we believe,
ultimately accept "mutual assured destruction" as an accurate description
of one facet of the current superpower relationship. "Assured destruction,"
however, is an extremely simple variant of deterrence. Contemporary U.S.
policy might better be described as invoking "extended protracted flexible
deterrence." This label evokes a much more complicated, spatially differ-
entiated, temporally denominated, map C - S - X on the U.S. side.
Equivalent complexity is often claimed for the Soviet threat.
Extensions of "Deterrence" in Theory and Practice

Much theoretical work on deterrence theory seeks to reconstruct the
idiosyncratic, unpredictable, and suboptimal components of the historical
practice of deterrence, expecting the reconstructed information product to
be cleanly structured and simply explained. As Robert Jervis has written,
"[l)f we can build generalizations about the motivated and unmotivated biases
that create deviations from rationality, we can probably use them to replace
the rationality postulate in deterrence theories while preserving the deductive
structure of the theories and so retaining the benefits of power and parsimony"
(1985: p. 11).
Jervis and his colleagues seek to build "transducers" that convert
information available to the parties in a deterrent situation into information
filtered and packaged in the form that they can assimilate and that helps
to account for their actions. We believe that designing "transducers" of this
sort must extend well beyond the removal of sources of bias in the parties'
perceptions. Indeed, it must extend to describing how the parties categorize

perceptions to conclude that they are witnessing any thing at all. But the
process of categorization is and must remain plastic. What John Kennedy
first "saw" in the aerial reconnaissance was not threatening; he had to be
instructed to "see" the threat. From that lesson, he inferred that the Soviets
were challenging the U.S. deterrent posture.
Were a U.S. president today to "see" similar photographic evidence, it
is questionable what interpretations he could be taught to make, and how
the interpretations would relate to the contemporary practice of deterrence.
lt is not easy to characterize contemporary U.S. posture, which minimally
involves protracted, extended, flexible deterrence with limited nuclear options.
An enemy could "challenge" numerous facets of this posture. Which would
constitute a "crisis"? What actions might be viewed as undermining the
stability of "deterrence"? Adequate answers cannot be produced by positing
appropriate perceptual filters and attention rules. The "official" meanings,
meanings-in-use, and extensibility of the concepts are all centrally involved.
To organize one's security around unintelligible strategic concepts invites
immoderate decisions about strategy.
Although the United States and USSR are re-creating detente at the
military level, the possibility of another superpower crisis must not be
underestimated. If a future situation is perceived by leaders as a "crisis,"
then it is very likely to involve apparently novel phenomena, whose cate-
gorization is highly problematic. How can the analyst predict whether or
not a hypothetical situation will be conceived as a "crisis of deterrence"?9
If leaders' pre-crisis conceptualizations of "deterrence" are ill-formed or
significantly incomplete, then the extensions of the concept they make in
stressful future contexts can be essentially open-ended. Concomitantly, the
action options they envision-and the very nature of the "decision" with
which they find themselves confronted-might be radical and radically
unpredictable. In other words, we cannot expect to comprehend decisions
until we understand how the decision context is framed, and to understand
the latter we must anticipate how participants in the decision context
categorize the information they receive. Current declaratory doctrines of
"deterrence" are among the most complicated political concepts with which
leaders must grapple. If these doctrines persist as doubtfully intelligible,
then the space of political possibilities will remain dangerously open despite
the progressive accretion of arms control and crisis management agreements.
Notes
1. More analytically inclined government officials apparently use deterrent slogans
imprecisely or vaguely; it is difficult to draw conclusions about the quality of reasoning
from political debates (Bennett 1987).
2. Problems of mutual knowledge can be treated as constructing accessibility

relations among environments. In the models of belief systems of strategic parties,
Wij is always accessible from Wi. Thus we indicate that the "outermost" environment
is passed as a parameter and assume that the parties, given this information, can
demarcate embedded environments describing their models of the beliefs about
others.
3. More than four distinct roles can be identified, and the actors that we have
described as individuals can be coalitions, either of states or of organizations within
a state apparatus. These extensions are explored in Bennett, in preparation.
4. More carefully stated, contingency C happened, if at all, sufficiently distantly
in the past that a regards its distinct reoccurrence as conceivable. lt cannot be the
case that C "is happening now" or that nonrepetitive C "just happened." More
concretely, one does not attempt to dissuade an enemy from attacking when the
attack is already under way, though one might try to dissuade the enemy from
continuing the attack (as distinct from beginning it) or escalating it. Searle and
Vanderveken (1985: eh. 2) integrate linguistic tense into speech act analysis. lt is
possible to model such situations by making deter recursive in C, described by a
list of successive contingencies ordered from least to most severe. Thus one can
explore requirements for effecting postures of "escalation dominance" and other
forms of "deterrence after deterrence has failed."
5. Our model of "deterrence" shows b to infer C' entirely from Ua to infer S'
from both Ua and S. This happens because deter? takes as arguments information
including Ua and S. The reader may wonder why b is not assumed to infer S' solely
from Ua. Nonlinguistic display of the sanction-e.g., exercising military capabilities-
can play a significant role in b's estimate of a's sanction. Further, one might model
certain parties b as estimating a's intentions (as they appear on other conditions
of threaten?) entirely from observed capabilities, while other parties make their
estimates chiefly from information drawn from a's expressed intentions in Ua. Indeed,
by exploring alternative specifications of this sort in predicate infer, one can examine
the implications for achieving varieties of stable deterrent situations under different
models of practical reasoning. Application of speech act theory emphasizes the
utilization of linguistically encoded information, which we have "packaged" in Ua.
But one cannot ignore the role of direct observation or experience, which, though
in quotation, is "packaged" also in S.
6. Because b places its estimate of the contingency C' in the context of b's
beliefs about a's beliefs, this says that b thinks that-via reference to Waba-a
doesn't think that the contingency has happened.
7. Minimal conditions for threaten? require both predicates prefer and intend.
Deterrent acts are inevitably embedded in background activities related to a, b, and
other parties. An overly simple model of practical reasoning-from wants or needs
to intentions and then to actions-derives intentions directly from preferences. But
more adequately realistic models of practical reasoning acknowledge vital mediating
roles for causal understandings, conflicting wants, material obstacles, and anticipated
reactions by others. Consequently, information about a party's preferences among
alternative actions and information about its intentions are both necessary ingredients
to reasoning about constituting different acts. Preferences cannot be deduced simply
from intentions; nor can intentions be expressed reductionistically simply as relations

among preferences.
This chapter stops short of making the next level of decomposition, which
interprets intentions as conclusions of practical "syllogisms." In other words, from
information about preferences (and other "causal" information) the parties infer
beliefs about intentions. Thus predicates intend and infer are intimately related. The
interface of intentionality and practical reasoning is treated in Bennett, in preparation,
eh. 5.
8. Matters are actually somewhat more complex. For simplicity we have assumed
that there is mutual agreement about the relevant texts Ua and Ub and-more
controversial still-about the sanctions (Sa and Sb) that each threatens to inflict
upon the other.
9. In Soviet perceptions, slightly different concepts are implicated, though the
principle is the same. Official Soviet writings have, except in periods of detente,
generally rendered the U.S. declaratory doctrine of deterrence as ustrashenie, which
can mean "compellence" or "coercion" as well as "deterrence." In contrast, they
have tended to label their own policy as sderzhivanie, which connotes retraining an
opponent and combines aspects of "defense" with (the Western concept of) "de-
terrence." Western analysts of Soviet strategic thought have tended to treat the
correspondence of terms in the English and Russian languages as problematic, but
not the intelligibility of the terms within their respective languages. Informative
exceptions to this preoccupation with term-for-term translation include Lambeth
(1981) and Ermarth (1981).
Bibliography
Abelson, Harold, Gerald Jay Sussman, with Julie Sussman (1985) Structure and
Interpretation of Computer Programs, MIT Press.
Alker, Hayward R., Jr. (1988) "Toward Generalized Event-Data Research on Avoidable
War," DDIR-Update 3:2, 2-4.
___ . (1984) "Historical Argumentation and Statistical Inference: Towards More
Appropriate Logics for Historical Research," Historical Methods 17:3, 164-173
and 270.
Alker, Hayward R., Jr., James P. Bennett, and Dwain Mefford (1980) "Generalized
Precedent Logics for Resolving Insecurity Dilemmas," International Interactions
7:165-206.
Austin, J. L. (1962 ) How to Do Things with Words, Harvard University Press.
Bennett, James P. (1984) "Data Stories: Learning about Learning from the U.S.
Experience in Vietnam," in Foreign Policy Decision-Making, Cognition and Artificial
Intelligence (D. Sylvan and S. Chan, eds.), Praeger, pp. 227-279.
_ _ . (1987) "Deterrence in Contemporary Policy Debates: The Expression of
Strategic Doctrines in Political Arguments about SDI," paper presented at the
annual meeting of the International Studies Association.
___ . (in preparation) Deterrence and Its Afterthoughts.
Brodie, Bernard, ed. (1946) The Absolute Weapon: Atomic Power and World Order,
Harcourt Brace and Company (1972 reprint, Books for Libraries).
Brodie, Bernard (1959) Strategy in the Missile Age, Princeton University Press.
Bundy, McGeorge (1988) Danger and Survival: Choices about the Bomb in the First
Fifty Years, Random House.
Ermarth, Fritz (1981) "Contrasts in American and Soviet Strategic Thought," in
Soviet Military Thinking (Darek Leebaert, ed.), Alien and Unwin, pp. 50-69.
Freedman, Lawrence (1981) The Evolution of Nuclear Strategy, St. Martin's Press.
Friedberg, Aaron L. (1984) "A History of the U.S. Strategic 'Doctrine'-1945-1980,"
in The Art and Practice of Military Strategy (G. E. Thibault, ed.), National Defense
University.
Garthoff, Raymond L. (1985) Detente and Confrontation: American-Soviet Relations
from Nixon to Reagan, Brookings.
___ . (1987) Reflections on the Cuban Missile Crisis, Brookings.
___ . (in press) Reflections on the Cuban Missile Crisis, rev. ed., Brookings.
Gray, Colin (1988) The Geopolitics of Superpower, University of Kentucky Press.
Haas, Andrew (1986) "A Syntactic Theory of Belief and Action," Artificial Intelligence
28:245-292.
Huth, Paul (1988) Extended Deterrence and the Prevention of War, Yale University
Press.
Huth, Paul and Bruce M. Russett (1984) "What Makes Deterrence Work? Cases from
1900 to 1980," World Politics 36:496-526.
___ . (1988) "Deterrence Failure and Crisis Escalation," International Studies
Quarterly 32:29-45.
Jervis, Robert (1979) "Deterrence Theory Revisited," World Politics 31:289-324.
Jervis, Robert, Richard Ned Lebow, and Janice Gross Stein, with Patrick M. Morgan
and Jack L. Snyder (1985) Psychology and Deterrence, Johns Hopkins University
Press.
Kaufmann, William W. (1954) The Requirement of Deterrence, Princeton University
Center for International Studies.
Lambeth, Benjamin (1981) "How to Think about Soviet Military Doctrine," Soviet
Strategy (J. Baylis and G. Segal, eds.), Croom Helm, pp. 112-116.
Levy, Jack S. (1989) "Quantitative Studies of Deterrence Success and Failure," in
Perspectives on Deterrence (P. C. Stern et al., eds.), Oxford University Press, pp.
98-133.
Mearsheimer, John (1983) Conventional Deterrence, Cornell University Press.
Morgan, Patrick (1977) Deterrence: A Conceptual Analysis, Sage Publications.
Rosenberg, David Alan (1987) "U.S. Nuclear Strategy: Theory versus Practice,"
Bulletin of the Atomic Scientist 43:2 (March), 20-26.
Russett, Bruce (1963) "The Calculus of Deterrence," Journal of Conflict Resolution
7:97-109.
Schelling, Thomas (1966) Arms and Influence, Yale University Press.
Schneider, Barry R. and Colin S. Gray (1984) Missiles for the Nineties: ICBMs and
Strategic Policy, Westview Press.
Searle, John R. (1969) Speech Acts: An Essay in the Philosophy of Language,
Cambridge University Press.
Searle, John R. and Daniel Venderveken (1985) Foundations of ll/ocutionary Logic,
Cambridge University Press.
Stern, Paul C., Robert Axelrod, Robert Jervis, and Roy Radner (1989) "Deterrence
in the Nuclear Age, the Search for Evidence," in Perspectives on Deterrence (P.
C. Stern et al., eds.), Oxford University Press, pp. 2-14.
Wolfers, Arnold (1962) Discord and Collaboration: Essays on International Politics,
Johns Hopkins University Press.
6
The Computational Modeling of
Strategic Time
Howard Tamashiro
Since its inception three decades ago, artificial intelligence has introduced
many computational tools for describing the representation and processing
of information. This chapter will suggest how computational ideas can shed
new light on a neglected area of political statecraft: temporal reasoning in
strategic contexts.
In general, AI/ computational modeling offers several distinct advantages
over other modeling approaches:
1. Because mental or social processes might be postulated to be com-

putational (i.e., thinking is represented as the rule-governed manipulating
of symbols), computational methods can be represented as strong
simulations. In this mode they seek to elucidate cognitive or social
structures, processing, and outcomes in contrast to weak simulations
such as probabilistic or regression models that seek to exhibit only
accurate outcomes.
2. Computational modeling postulates an active, creative cognitive, or
social capability rather than a predominately passive/adaptive infor-
mation-processing makeup. Hence, counterfactual reasoning and al-
ternative policy actions can be investigated.
3. Computational models can handle both numeric and symbolic data
using the varied, informal processing and representation procedures
found in human thought. This means a wider array of knowledge
claims and richer models can be investigated. In contrast, non-At-
inspired modeling tends to focus on a narrower set of input structures
(e.g., digits, events data) and a narrower sort of processing (e.g.,
deduction, statistical manipulation).
4. Computational modeling expands our ability to examine how ontological
assumptions influence our understandings. Such fundamental intelli-
149
150 Howard Tamashiro
gibility issues are largely intractable via non-AI-based models and,

consequently, have been heretofore neglected.
In what follows I hope to show that a computational approach offers

ideas for representing and using time that surpass traditional, noncompu-
tational attitudes. While a working AI program will not be exhibited, the
ideas to be explored here are derived from the AI/ computational literature
and illustrate many of the claims enumerated above.
The Emergence and Implications of "Subjective" Time

Time is a basic metaphysical feature that underpins higher cognitive and
perceptual activities. In most traditional and behavioral political science
research, time is treated as a background metric for describing event
changes. This approach implicitly assigns time both a descriptively neutral
role and great ontological simplicity. Unfortunately, this view makes it difficult
to include time as a perceptual variable since time sense embodies phenomenal
aspects and is not a purely objective-neutral feature (Fraisse, 1984).
Cognitively based AI methods do not suffer this restriction. They can
model such "subjective" time-based activities as intentional-inferencing
(Thorson, 1984), context sensitive reasoning (Barr and Feigenbaum, 1981),
counterfactual inferencing (Thorson and Sylvan, 1982), temporal inferencing
(Alien, 1985), and automatic planning (McDermott, 1982). Currently, time
can be conveyed in direct situational descriptions (as in "Diplomatic relations
were severed yesterday"), in propositional attitudes (as in "I believe diplomatic
relations were severed yesterday"), or in compliance conditions (e.g., com-
mands, task assignments). Hence, we have now the means not only for
describing event changes but also the emerging means for describing
perceptual, time sense changes. The next, logical step would be to model
temporal maneuverings for strategic advantage.
An implication of the computational-strategic view of modeling time is
worth elaboration. Computational processes are driven by the structural or
syntactic properties of the representations being processed. Computations
are insensitive to the content of semantic properties of representations such
as truth, reference, and meaning; these semantic considerations exist on
the representational level. Hence, our political theorizing is substantively
expressed on the semantic level, but the necessity of constructing coex-
tensively a running program on the syntactic-computational level forces us
to construct exceptionally disciplined and detailed conceptualizations. In
particular, this semantic/syntactic bifurcation allows computationally rich
descriptions of time, while permitting sensitivity to the semantic implications
of our structural assumptions. Belief-intentional content can be married with
computational experimentation.
Computational Modeling of Strategic Time 151
This computational opportunity is very attractive because many politically

important temporal experiences are not reducible to neutral metric or indexical
descriptions (e.g., clock time, calendar time). For example, during the SALT-
11 debate, conservatives attacked the treaty provision that allowed the Soviets
thirty Backfire bombers a year. However, in publicity campaigns this provision
was described as allowing the Soviets to build a "new Backfire bomber
every twelve days." One can give an AI account along the following lines:
The treaty provision was purposely presented in a temporal form that invited
people to encode it in a way that implied "the Soviets are allowed very
many Backfires." This temporal maneuver assumes people's processing
capabilities are limited so that few will expend the effort to realize that "one
Backfire every twelve days" equals "thirty Backfires a year." By this ploy
people were invited to construct subjective temporal maps very different
from a neutral, "calendar" understanding.
This SALT-11 example is indicative. In many foreign policy contexts,
statements that seem to be "behavior descriptions" (e.g., "under SALT-11
one Backfire will be allowed every twelve days") are actually intentional and
more strongly, strategic propositions (Thorson and Andersen, 1987). Time
conceptions must be constructed with a sensitivity to this strategic dimension.
The "date line" view of time, which suffices for purely descriptive aspirations,
will not do. Computational modeling offers an important option here. Unlike
metrical-indexical accounts of time that rely on syntax (digital-structural
properties) alone, computational modeling can reflect both semantic and
syntactical considerations. This allows the modeling of both numerical and
cognitive features of time sense.
Computational modeling offers theorists another investigative possibility
heretofore largely inaccessible to other modeling techniques. lt may allow
the study of ontological influences on political thought. What counts as
something being politically intelligible depends to a great degree on subtle,
poorly understood ontological commitments (Thorson, 1987).
The importance of ontological assumptions is perhaps best clarified by
the following puzzle (after Putnam, 1988): Suppose you take someone into
a room containing a lamp, notebook, and pen resting on a table, and nothing
else, and ask "How many objects are in the room?" If one counts only the
lamp, notebook, pen, and table, then one might answer "four." But, what
of the two people standing in the room? In that case, one might answer
"six." But, what of the pages in the notebook? Or the atoms in the room?
Or the collection consisting of the lamp and the noses of the two people
in the room? Is that last collection an "object" to be counted? The point
of the puzzle is that the same situation can be described in many different
ways, depending on our choice of conceptual scheme. The situation itself
does not identify the "one true description" or the "facts" of the situation,
or the correct meanings of words like "object" and "exist." These things
are legislated for us when we adopt ontological commitments. Lest one

dismiss this problem (called in philosophy the problem of "conceptual
relativity") as hopelessly removed from real-world politics, consider the
following example.
In the opening stages of the Vietnam War, when the United States decided
to communicate its resolve to Hanoi via bombing, key U.S. leaders showed
little curiosity about the target list (Gibson, 1986). They were interested
simply in generating "bombing events." But, the "event-content" of the
bombings (destruction, casualties, etc.) had important political implications
for Hanoi. It is unlikely North Vietnam saw the strikes merely as "bombing
events." If the United States was to be successful in making its "political
message" intelligible to Hanoi, both countries had to share similar ontological
assumptions. The United States displayed little awareness of this, which is
surprising because in its nuclear strategy, target distinctions for the purposes
of signaling were much studied. In short, ontological differences existed not
only between the United States and North Vietnam, these differences also
existed within the U.S. government over various conflict scenarios (e.g.,
Third World limited war vs. superpower nuclear war). .--
Computational methods may prove useful for tracing some of these
ontological dependencies and studying how ontological commitments are
extended to encompass new situations (see the Thorson and Bennett piece
in this volume). These computational possibilities have immediate relevance
for studying time. Instead of seeing time as a metrical variable, it is more
useful in strategic contexts to conceive of time as a framework of varied
organizing (ontological) assumptions.
Within the realm of ontological possibility, some temporal outlooks are
apt to be more theoretically interesting than others. We should not, however,
expect a single all-purpose model of time to emerge, as different time senses
are useful for different purposes. A computational modeling perspective
suggests the following guidelines for selection:
1. We want time conceptions exhibiting features that are computationally

tractable. Some rudimentary syntactical framework is needed (this
would omit mystical or kinesthetic understandings of time). We also
wish an accompanying semantic (interpretive) framework so the tem-
poral outlook can be talked about and justified on politically substantive
grounds.
2. We want time conceptions that are consistent with known human
cognitive capabilities. A key AI theme is that human information
processing limitations are significant; i.e., people are "satisficers" rather
than "maximizers." Accordingly, we should pay close attention to time
conceptions consistent with satisficing and the conserving of com-
putational resources.
3. Since time strategies are our main interest here, we want time
conceptions that allow for the study of interesting time maneuverings.
We now turn to the examining of concrete examples to identify interesting

temporal features and strategies.
Some Interesting Temporal Features and Strategies

AI research indicates that what constitutes knowledge and how it should
be represented is not always clear (Reddy, 1989), and this is certainly the
case with time. A long-standing barrier to its strategic study is its perceptual/
contextual complexity. Time perception can be biased in many different
ways. One path around this problem is to identify how time has been used
strategically in politics and then note, in a preliminary fashion, those temporal
elements that were manipulated. In this way, features might be identified
for later modeling efforts.
Events
To speak of time means speaking of events. Psychologically, we are
aware of time because of changing events; events are perceivable, time is
not.
Taking time sense to be event-based does not imply that individual events
are recalled. Memory limitations may preclude this. For example, the Chinese
feeling of greatness does not depend on recalling individual triumphs, but
on a generalized sense of continuity spanning 3,000 years (Pye, 1984).
Assuming that time's meaning derives from some description over events
introduces the issue of conceptual relativity discussed earlier. Since people's
rules for selecting and defining events differ, one event set for interpreting
time could easily disagree with another, hence producing divergent under-
standings and possibilities for strategic behavior. Imposing a single, preferred
historical account while marginalizing others is one sort of strategic move
designed to control time sense.
With regard to strictly syntactical matters, we note that AI has developed
point-based (McCarthy and Hayes, 1981), interval-based (Alien, 1985), and
mixed systems (Shoham, 1988) for temporal representation. All these
approaches are consistent with an event interpretation of time.
Event Sequences
Event ordering is fundamental for most temporal outlooks. Ordering helps
create an intelligible background context and makes causal reasoning possible.
Accordingly, the manipulation of event ordering is an important temporal
strategy when trying to establish a preferred event description. An ordinary
example would be the back-dating of a check to avoid a late payment

penalty. Similar event ordering ploys are used for political purposes. For
example, at a crucial point in the 1962 Cuban missile crisis, Washington
received two successive proposals from Moscow-the first was conciliatory;
the second, hard-line. Washington in effect transposed the message sequence
by choosing to regard the first proposal as the most recent and settled the
crisis on that basis. During Kissinger's 1973 shuttle diplomacy efforts to
mediate between Egyptian President Sadat and the Israelis, Kissinger got
at times both Sadat's bargaining and fallback positions, and then released
these in a sequence designed to induce similar concessions from Israel
(Golan, 1976; Licklider, 1988). Most recently, a major clash between Rockwell
International Corporation and the Defense Department erupted over an "event
ordering" disagreement. After having discovered fraudulent double-billing of
the Pentagon by some of its employees, Rockwell voluntarily informed the
government and offered restitution. To Rockwell's surprise, the government
responded by slapping Rockwell with criminal indictments. According to
the government, the sequence of events in Rockwell's story is wrong. "The
company didn't even mention voluntary disclosure until the U.S. attorney's
office mentioned indictment" (Read, 1989).
This potential for switching the order and, hence, the meaning of events,
is not limitless. In political contexts, expectations, conventions, organizational
procedures, resource limits, and so on can impose strict boundary conditions.
Nonetheless, opportunities to mix the temporal sequence in event descriptions
do emerge with surprising frequency, especially in disputes over events
leading to the onset of crises or wars.
Events Insertion and Deletion

Another ploy is to delete events from some account in order to deceive
or shield-or as little Joey explained to his mother, "The whole thing began
when Billy hit me back." Overlooking events, for example, is a common
rule governing covert action designed to give a president "plausible deniability"
(Treverton, 1987).
Inserting events into an event stream, commonly referred to as the
"timing" of events, is also a standard diplomatic maneuver (e.g., the fait
accompli). In general, the ability to insert events while denying others this
option is an important form of power. Israel, for instance, declared inde-
pendence in 1948 ten minutes after the end of the British mandate in order
to preempt any U.S. move toward a UN trusteeship option (Yaniv and Lieber,
1983). More interesting is the timing of threats to minimize the risk they
will have to be acted on, while maximizing the political payoffs. The Soviets
seem especially inclined toward this strategy. In both the 1967 and 1973
Arab-Israeli wars, they threatened intervention against Israel only during the
final stages of hostilities when a military resolution had already appeared.

"Threats that cannot fail" were also issued by Moscow during the 1956
Suez Crisis and the 1961 Bay of Pigs operation (Cohen, 1981).
Establishing Past, Present, and Future Distinctions

Past, present, and future distinctions are mandatory for studying temporal
politics. Besides being deeply rooted in the political discourse of policy-
makers, the fundamental notions of planning and strategy are inextricably
tied to future expectations, past behaviors, and present actions. Even though
the present separates the past from the future, the duration of the present
can vary. For example, the Watergate scandal and Teheran hostage crisis
remained part of Nixon's and Carter's "political present" much longer than
the Iran-Contra affair did for Reagan. In short, past, present, and future
distinctions are open to strategic manipulation.
To escape from past embarrassments, one can use salient events (e.g.,
elections, cabinet shakeups, changing ambassadors, etc.) to demarcate and
discount the past while shifting attention to the present or future. For
example, in the early 1960s, the U.S. military kept insisting, incorrectly,
that the Vietnam War was progressing well. But, immediately after the
overthrow of South Vietnam's President Diem, a flood of truthful reporting,
heretofore suppressed, revealed the Communists had been winning. The
military position immediately changed to accept the negative reports but
claimed that everything had been going well until the coup (Halberstam,
1973). Diem's fall was a convenient event used by the Pentagon to shift an
awkward policy position into the "past." The 1976 appointment of a new
CIA director by Carter was used in a similar way to break with past scandals,
as was the 1987 appointment of NSC adviser Carlucci following the lran-
Contra affair.
In certain institutional contexts, past, present, and future distinctions can
be simply stipulated. A particular instance is Reagan's controversial retroactive
hiring freeze, initially planned to take effect from November 1980, two
months before the 1981 inauguration. This move was designed to nullify
prior Carter appointments. Retroactive budget cuts are another device for
organizing time to suit political purposes.
Time Horizons
Closely related to past and future distinctions is the notion of time horizon
or temporal perspective. Time horizon involves more than a collection of
events; some relation or causal connection of interest must exist over those
events. Computationally, horizons are focus-of-attention, pattern-recognition,
and complexity-regulating devices. Politically, horizons are necessary for
priority setting, forecasting, planning, and discounting. Short time horizons
encourage tactical, present-centered perspectives, where complexity is re-

duced (Simon, 1981), but the impact of transitory matters and the weight
of unforeseen events tend to be magnified. Conversely, long time horizons
encourage strategic perspectives where the importance of transitory per-
turbations and the pressures for rapid reaction are relaxed. However, as the
time horizon expands, it becomes more difficult to falsify predictions-or
as an old Trotskyite once boasted, "Proof of Trotsky's farsightedness is
that none of his predictions have come true yet." In situations where the
relevant time horizon is unclear, as in protracted nuclear war, planning
becomes very difficult if not impossible.
The subjective nature of time horizons makes it a useful political tool.
In particular, extending one's temporal perspective tends to decrease the
saliency of events. For example, European Communist leaders now discount
the East-West conflict by arguing the historical race between capitalism and
communism is open-ended; hence, Marxist goals are not critical and may
even be unrealizable (Markham, 1989). On the other hand, reducing one's
time horizon tends to sharpen conflict. Foreign policy critics, for instance,
tend toward short-range perspectives insisting on relatively quick, continuous
policy results (Holsti, Siverson, and George, 1980).
Duration
Time duration has been the most extensively studied of all temporal
features. Psychological research suggests that time estimation is extremely
context sensitive and can be biased in very many different ways depending
on whether one is estimating elapsed time alone or the duration of a task,
whether one's expectations have been satisfied or not, whether one's ex-
pectations are pleasant or not, whether the setting is complex or not, and
so on (Fraisse, 1984). Political efforts emphasizing duration seem especially
common in critiques of ineffective policies.
Markedly new methods of representing political duration are also appearing.
These representations regard time intervals as durationally equivalent because
of similarities in their event compositions even though their calendar-measured
durations differ. So, two tabor strikes of unequal calendar duration may be
regarded as durationally equivalent with respect to man-hours of productivity
or profits lost. Another example is the Singer and Small (1968) claim that
"the 85 years prior to 1900 represent an approximate equivalent of the 45
post-century years in terms of 'diplomatic time.' " Here, the notion of
equivalent "diplomatic time" depends on a special, context-determined set
of events (wars and alliance commitments) that departs from and takes
precedence over calendar time. Zinnes (1983) has mathematically formalized
this irregular, event-driven outlook where event sequences are taken as the
measurement standard. Alien (1987) has proposed still other nontraditional
"social clock" representations of time. All these systems can be compu-

tationally modeled.
Synchronization
Synchronization is not a simple concept. A first-cut characterization might
link synchronizing with rate coordinating or placing temporal cycles in
phase. But ambiguity remains because this might mean (in ascending order
of difficulty) simultaneous actions, sequential actions, or controlled rates of
actions. These definitional difficulties aside, however, the notion of syn-
chronization (in all forms) seems to be an important dimension of politics.
The imposition of preferred time cycles by dominant groups over weaker
ones, for instance, seems common in certain parliamentary contexts (DeRidder
and Tamashiro, 1982) and in international affairs (Gieditsch, 1974).
Consider the possibilities of strategic synchronization in dialogues. For a
dialogue to be understood, it must be placed in some context-this context,
in part, is determined by the synchronized exchange of messages between
the conferees. If this message sequence is jumbled, then confusion will
result. For example, during the latter stages of the Cuban missile crisis, as
the pressure and tempo of events increased, synchronization in the dialogue
between Washington and Moscow began to break down so that each side
found it harder to decipher the intentions of the other. One major sequential
confusion, still unresolved today, concerns Khrushchev's two letters to
Kennedy at the height of the crisis. The first letter, received on Friday
evening, October 26, was conciliatory and offered promise of a peaceful
settlement. The second letter, received the next morning, was hard-line and
threatening. Understandably, this discrepancy caused much anxiety and
confusion in Washington.
Analysts have since offered different explanations: (1) the "soft" Friday
letter was Moscow's real offer and the "hard-line" letter was an outdated
demand released by mistake, (2) the "hawks" in Moscow insisted both
letters be sent, (3) Khrushchev's "soft" letter was a personal initiative that
was later overruled by the "hawks," and the "hard-line" letter was the real
Soviet position.
Alexander George (1971) has suggested another explanation, one that
assumes the ambiguity in message synchronizing was intended as a strategic
maneuver by Moscow and was not the result of mistakes or internal
disagreements. The Soviets sent the first, "soft" letter to calm escalating
pressures in Washington. The second, "hard-line" letter was the ploy of
"raising the price after an opponent has agreed to your first proposition,
in effect trying to get him to pay twice" for the same concession. If George
is correct, then this Soviet jumbling of messages was a striking example
of temporal politics-the use of synchronizing ambiguities to test U.S. will
and raise the ante without risking the dangers of an irreversible "take-it-
or-leave-it" ultimatum.
Two Preliminary Observations

Two important insights emerge from the previous survey, which was not
meant to exhaust all temporal events of potential interest. First, time strategies
"work" because we are reconstructive as well as constructive creatures. We
can alter time sense, both for ourselves and others, by the way we recall
the past. Time strategies are designed to influence, in different ways and
directions, that recalling and the judgments flowing from that recalling.
Further, because these reconstructive effects require semantically rich con-
texts, the computational approach is the only one currently capable of
systematically modeling these strategies and examining how they produce
different, contending political interpretations. I
The second observation concerns the inferencing mode appropriate for
investigating reconstructive properties. Certain types of reasoning (e.g.,
mathematics) are monotonic in that old conclusions remain valid even with
the addition of new facts. But such an axiomatized, mathematical logic is
not possible for temporally based, reconstructive reasoning (Shoham, 1988)
because reconstructive reasoning is nonmonotonic. Nonmonotonic reasoning
"jumps to conclusions" by inferring things both from what is known and
what is not known; it may drop old conclusions as new facts emerge; and
it may re-embrace earlier discarded conclusions. For example, if you know
politicians are worried about deficits, you may conclude that taxes will be
raised. If you then learn it is an election year, you may conclude taxes will
not be raised. And finally, if you learn the elections have been postponed,
you may again conclude taxes will be raised. As more premises are added
(and the relevant time horizon changes), the conclusion keeps changing.
Nonmonotonicity in reconstructive reasoning means inferences are tentative,
time dependent, and potentially reversible.
If complete axiomization is impossible and semantics is primary, then
what is the alternative? While no definitive answer exists, work is proceeding
on methods of formalizing nonmonotonic reasoning. Some elements of these
efforts are outlined below, together with their implications for modeling time
strategies.
Formal Considerations and the

Inherent Uncertainty of Strategic Environments
Any computer implementation requires formalization. Logic in AI is a
theoretical description language used to formally represent computer program
capabilities. To the extent that At-based theories are so expressed, the
greater will be the foundational clarity of one's theorizing. Logical arguments

are also useful if they strengthen our intuitive understanding of some
theoretically important problem. Recent work on formalizing temporal rea-
soning makes such contributions.
To effectively capture the nonmonotonic nature of temporal inferencing,
nonmonotonic modal logics have been proposed (Shoham, 1988). In order
to convey an intuitive feel for nonmonotonic modal logic, we first characterize
nonmonotonic logic by contrasting it with standard logic. Later a brief
characterization of modal logic will be presented.
In standard or classical logic, the point in interpreting a formal system
is not to determine what logical formulas are true under some particular
interpretation. The goal rather is to determine what formulas are true under
all interpretations of that system (an interpretation under which a set of
formulas becomes true is said to be a model for those formulas). Showing
a given formula is true under all interpretations or models demonstrates
that formula is a "truth of logic" with respect to that formal system.
But in nonmonotonic logic, as noted above, a logical formula may not
be true for all models-i.e., old inferences maybe reversed with the addition
of new facts. Standard or classical logic cannot handle such "contingent
inferencing." In contrast, nonmonotonic logic takes a logical formula to be
true with respect to "preferred" models, which are a subset of all possible
models. In this way, different preference criteria can produce different
nonmonotonic logics. For example, the preference criterion might be time-
based, if one is interested in temporal reasoning. Thus, different things can
be accepted as true, depending on one's preferred temporal outlook.
A second branch of logical exploration exists, distinct from nonmonotonic
logic, called modal logic, which also promises insights into temporal reasoning.
In classical logic, truth-functional operators (e.g., or, and, if . . . then)
operate on statements to form new statements. The truth-value of the new
statement formed is determined uniquely by the truth-value of the component
statements. Moreover, truth-functional operators combine statements to
produce indicative claims about what is the case, not about what must be,
or could be, or is thought to be.
In contrast, modal logic relaxes these restrictions. Modal logic allows
discussion of moods other than the indicative (e.g., what must be, might
be, will be, and so on). Further, modal operators form statements whose
truth-value is independent of the truth-value of the component statements.
This can occur because modal statements are interpreted against many
"possible worlds" (technically referred to as a Kripke structure). A possible
world may be the actual world or it may be a nonactual but logically possible
world. So, a modal statement might be true in one possible world and false
in another (e.g., "It is possible the current year is 1989").
The flexibility of modal logic is useful on a number of counts for studying

temporal reasoning. First, modal systems can accommodate past, present,
and future distinctions, or more complicated time variable treatments (Snyder,
1971). Second, modal systems can formalize conditions of our knowledge
of truth and falsity, instead of simply the conditions of truth and falsity,
as is the case with classical logic. Third, modal formalization can clarify
certain implications of our theorizing. These points will be illustrated below.
While modal logic is monotonic, nonmonotonic versions have been
formulated for studying temporal reasoning. One version employs modal
logic as a logic of knowledge (Shoham, 1988). In this application of modal
logic, the set of all possible worlds is the set of all the ways the world
could have been, compatible with what an agent knows (which is not
necessarily what "is") in the given world. These modal characterizations
allow us to distinguish three different sorts of temporal environments with
respect to the possibilities of strategic time: naturally stable time environ-
ments, parametrically stable environments, and strategic environments.
As a running example, consider a situation (after Shoham, 1988) where
a conservative-led, parliamentary government is expecting a year-end deficit,
and we wish to specify a reasonable inference rule that will allow us to
conclude when a tax hike is imminent. One formalization might resemble
the following:
1. O[(year-end government deficit) 1\ government is

fiscally conservative)/\
(no upcoming election)/\
(government budgeting based on a 1-year time horizon)/\
(tax collection system works)/\
. . . other mundane conditions
Implies
(tax hike))
This formalization assumes a "naturally stable" temporal context. The

time maneuverings described in Section 11 are not recognized possibilities.
The 0 -operator in front of the sentence means that the inference holds in
all possible worlds and takes on a law-like, causal quality. In this very orderly
context, time serves as a neutral, background relation.
An alternative formulation might resemble the following:
2. D(year-end government deficit) 1\ O(government is fiscally

conservative)/\
(no upcoming election)/\

(government budgeting based on a 1-year time horizon)/\
(tax collection system works)/\
. . . other mundane conditions
Implies
D(tax hike)]
This formulation departs from formulation 1 in that nonmonotonic results
are allowed. In order for the "tax hike" prediction to emerge, the D-
conditions must be known and justified explicitly. But, the 0 conjuncts act
as "boundary conditions" or default assumptions, which need not be stated
explicitly. So, the conjunct 0 (no upcoming election) embodies the "silent"
assumption that the scenario being described does not include an election.
The "tax hike" inference depends only on not knowing the negation of the
0-conditions. Hence, if one should learn of an upcoming election, this new
knowledge would cause the "tax hike" prediction to be withdrawn, thereby
producing a nonmonotonic reasoning pattern.
Formalization 2 offers the advantage of computational economy; only the
D-conditions need be specified in order to predict a tax hike. The 0-
conditions are automatically assumed. In contrast, for monotonic logics all
the antecedent conditions (e.g., no upcoming election, year-end deficit, etc.)
would have to be specified or else nothing could be inferred about a tax
hike. But a price is paid for the second formalization's computational economy.
By definition, the 0 conjuncts may not hold over every possible world that
is temporally accessible from the agent's given world. If we do not specify
explicitly when a 0 conjunct no longer holds, we may end up erroneously
predicting a tax hike in some future world. In short, as noted earlier on
nonmonotonic logics, we must define some notion of preferred models or
worlds that will exclude undesired knowledge models with unstated, abnormal
interruptions (e.g., where the tax collection system has collapsed, where a
nuclear war has destroyed the country, etc.).
One way of identifying preferred models where implicitly accepting the
0-conditions will not lead to error is to select only those models where any
knowledge changes in the 0-conditions occur as late as possible relative to
the tax hike. This notion of "late occurrence" can be rigorously defined
(technically called "chronological minimization"; Shoham, 1988) and has the
effect of selecting only knowledge models in which as "little as possible
happens." These preferred "chronologically minimal" models basically omit
abnormal interruptions. Further, if an abnormal occurrence becomes known
and explicitly accompanies formalization 2 as an axiom, then no chrono-
logically minimal model is satisfied, and hence no tax hike inference will
result.
Clearly, chronological minimization limits the knowledge needed to make
inferences (only the 0-conditions need be specified) while allowing for
nonmonotonic changes. But one must assume minimal or no knowledge
changes within certain time parameters-what might be called a para-
metrically stable environment. Fortunately, the nonmonotonic modal rep-
resentation of formulation 2 can accommodate broader time senses than
parametric stability. I say "fortunately" because inferring a tax hike clearly
falls within the realm of political intentions, and hence will be subject to
strategic perturbations. For example, instead of biasing our models by
preferring chronologically limited knowledge models, one might bias them
toward temporal maneuverings. If we expected political opponents, for
instance, to try inserting an "election" event into the time stream, we might
change the conjunct 0 (no upcoming election) to 0 (no upcoming election).
This change would bias formulation 2 so that we would never predict tax
hikes with elections. We could fail, however, to predict a tax hike if we
neglected to state explicitly that no elections were forthcoming. The 0 to
0 change, then, assumes that inserting elections to block tax hikes is a
frequent ploy. But, if such ploys are rare, we are better off with the original
formulation. Another possibility is to expand the budgeting time horizon in
the second 0 conjunct under the assumption that the government will
succeed in extending temporal perspectives, hence decreasing the saliency
and political opposition surrounding a tax hike. A still more radical change
is to rewrite the inference rule to predict O(no tax hikes), and then specify
all opposition time strategies that are likely to block such hikes. All such
changes would reflect a strategic time sense, in contrast to parametrically
or naturally stable ones.
One final observation emerges from these formalizing efforts. It is a
standard AI dictum that knowledge reduces uncertainty. However, this needs
to be qualified in strategic contexts where knowledge of time strategies can
be used to strengthen nonmonotonic properties, and hence inferential un-
certainty.
Conclusion
A new research framework requires its own vocabulary and outlook to
focus on fresh problems and details. We have explored some new concepts
and issues that At-based, computational modeling raises regarding strategically
directed temporal reasoning. The potential of computational methods suggests
the possibility of extending time sense beyond clock/calendar metrics to
encompass subjective temporal understandings. Among the important ob-
servations flowing from such an enterprise is the primacy of semantic
properties over syntax when representing temporal features. In other words,

modeling time sense is fundamentally an interpretive process, not one of
calculation. Another important observation is the need to capture nonmon-
otonic properties, since temporal reasoning is reversible.
Next, we explored how nonmonotonic modal logic might express temporal
reasoning. Besides modal logic, another alternative approach being inves-
tigated in AI is tense logic where tensed verbs are allowed in one's truth
claims so that sentences about the future may change their truth-values
with the passage of time. Another method is multivalued logic, where
sentences about the future are considered neither true nor false, although
they may become so when time elapses and new events emerge. Still
another, albeit nonformal, direction is the identifying of heuristics that guide
temporal inferencing.
Whether these efforts will succeed in capturing time strategies in all their
complexity is unknown. An important obstacle is computational tractability;
strategic reasoning usually leads to a combinatorial explosion of possibilities
that can quickly exhaust computational resources.
But questions of final design aside, attempts at developing computational
interpretations of time have already produced important conceptual contri-
butions. AI researchers are now more sensitive to the properties and features
of time sense in promoting political intelligibility and inferencing, and,
accordingly, they are developing more sophisticated representation schemes
for expressing temporal features in a computationally tractable form. In
short, temporal reasoning, in both its theoretical and applied forms, is a
burgeoning research area and likely to grow more so in the future.
References
Allan, P. "Social Time," in Cioffi-Revilla, C., Merritt, R., and Zinnes, D. (eds.),
Communication and Interaction in Global Politics. Beverly Hills, California: Sage,
1987.
Alien, J. "Maintaining Knowledge about Temporal Intervals," in Brachman, R. and
Levesque, H. (eds.), Readings in Knowledge Representation. Los Altos: Morgan ,
Kaufmann, 1985.
Barr, A., and Feigenbaum, E. Handbook of Artificial Intelligence, vol. 1. Los Altos:
Morgan Kaufmann, 1981.
Cohen, R. International Politics: The Rules of the Game. New York: Longman, 1981.
DeRidder, M., and Tamashiro, H. "Time Strategies in Coalition Crisis Bargaining,"
paper delivered at the Annual Meeting of the International Studies Association,
Cincinnati, Ohio, 24-27 March 1982.
Fodor, J. Representations. Cambridge, Massachusetts: MIT Press, 1981.
Fraisse, P. "Perception and Estimation of Time," Annual Review of Psychology, 35:
1-36, 1984.
George, A., Hall, D., and Simons, W. The Limits of Coercive Diplomacy: Laos, Cuba,
Vietnam. Boston: Little, Brown, 1971.
Gibson, J. The Perfect War. New York: Vintage, 1986.
Gleditsch, N. "Time Differences and International Interaction," Cooperation and
Conflict, 9(1): 35-51, 1974.
Golan, M. The Secret Conversations of Henry Kissinger. New York: Quadrangle,
1976.
Halberstam, D. The Best and the Brightest. Greenwich, Connecticut: Fawcett, 1973.
Holsti, 0., Siverson, R., and George, A. Change in the International System. Boulder,
Colorado: Westview, 1980.
Keegan, J. "The Human Face of Deterrence," International Security, 6(1): 136-151,
1981.
Licklider, R. Political Power and the Arab Oil Weapon. Berkeley, California: University
of California Press, 1988.
Markham, J. "Across a Divided Europe, An ldealogy Under Siege," New York Times,
p. All, 23 January 1989.
McCarthy, J., and Haynes, P. "Some Philosophical Problems from the Standpoint
of Artificial Intelligence," in Webber, Bonnie Lynn and Nilsson, Nits J. (eds.),
Readings in Artificial Intelligence. Palo Alto, California: Tioga Publishers, 1981.
McDermott, D.V. "A Temporal Logic for Reasoning about Processes and Plans,"
Cognitive Science, 6: 101-155, 1982.
Newell, A., and Simon, H. "Computer Science as Empirical Inquiry: Symbols and
Search," Communications of the Association for Computing Machinery, 19(3):
113-126, 1976.
Putnam, H. Representation and Reality. Cambridge, Massachusetts: MIT Press, 1988.
Pye, L. China: An Introduction (3rd ed.). Boston: Little, Brown, 1984.
Read, E. "Will Rockwell Case Hamper Self-Policing in Defense Industry?" Wall Street
Journal, p. At, 23 January 1989.
Reddy, R. "Foundations and Grand Challenges of Artificial Intelligence," AI Magazine,
9(4): 9-21, 1988.
Shoham, V. Reasoning about Change. Cambridge, Massachusetts: MIT Press, 1988.
Simon, H. The Sciences of the Artificial (2nd ed.). Cambridge, Massachusetts: MIT
Press, 1981.
Singer, J., and Small, M. "Alliance Aggregation and the Onset of War, 1815-1945,"
in Singer, J. (ed.), Quantitative International Politics: lnsights and Evidence. New
York: Free Press, 1968.
Snyder, D. Modal Logic and Its Applications. New York: Van Nostrand Reinhold,
1971.
Thorson, S. "Intentional lnferencing in Foreign Policy: An AI Approach," in Sylvan,
D., and Chan, S. (eds.), Foreign Policy Decision Making. New York: Praeger, 1984.
Thorson, S. "Working Memorandum Prepared for the Psychological Approaches to
Foreign Policy Conference," Maxwell School, Syracuse University, 8-9 May 1987.
Thorson, S., and Andersen, K. "Computational Models, Expert Systems, and Foreign
Policy," in Artificial Intelligence and National Security. Lexington, Massachusetts:
Lexington Books, 1987.
Thorson, S., and Sylvan, D.A. "Counterfactuals and the Cuban Missile Crisis,"
International Studies Quarterly, 26: 539-571, 1982.
Treverton, G. Covert Action. New York: Basic, 1987.

Yaniv, A., and Lieber, R. "Personal Whim or Strategic Imperative? The Israeli Invasion
of Lebanon," International Security, 8(2): 117-142, 1983.
Zinnes, D. "An Event Model of Conflict Interaction," in Zinnes, D. (ed.), Conflict
Processes and the Breakdown of International Systems. Denver, Colorado: University
of Denver, Monograph Series in World Affairs, 1983.
PART TWO
AI/IR Research:
International Events and
Foreign Policy Decision Making
7
Pattern Recognition of International Event
Sequences: A Machine Learning Approach
Philip A. Schrodt
Human political behavior is patterned: Certain sequences of behavior

occur on multiple occasions under predictable circumstances. The predict-
ability is stochastic rather than deterministic-the randomness of the pre-
diction exceeds the measurement error of the instrument-but political
behavior is far from totally random. For the past three decades scholars
of international affairs have been searching for statistical patterns, usually
employing variations on the normal distribution and the general linear model.
Computational models and other artificial intelligence techniques provide a
rich set of alternative definitions of pattern that may prove more effective
in modeling international behavior than the statistical patterns studied to
date.
The essence of pattern is repetition or regularity: Something is "patterned"
when one can infer unknown information about an object on the basis of
partial information about it. The term "plaid shirt" is a pattern that matches
a large number of objects that differ in calor, weight, size, and so forth,
but it excludes a much larger class of objects: "Plaid shirt" will not match
a car, a tie, the planet Mars, or even a striped shirt. These regularities are
not statistical, but they are patterns.
While Margolis's recent assertion that "pattern recognition is all there is
to cognition" (1987: 3) is stating the case excessively, pattern recognition
is a useful foundation for a computational model of politics (Schrodt
1984, 1985). The literature on pattern recognition in human problem solving
goes back to the 1960s (for example, Newell and Simon 1972; Simon 1982;
The Behavioral Correlates of War data utilized in this chapter were originally collected by
Russell J. Leng and were made available by the Inter-university Consortium for Political and
Social Research. Neither the original collector nor the consortium bear any responsibility for
the analyses or interpretations presented here.
169
Margolis 1987); a substantial literature on pattern recognition exists in the

AI literature 1 (for example, Fu 1974, 1982; Patrick and Fatu 1986; Devijver
and Kittler, 1987); another overlapping literature exists in statistics; and an
assortment of techniques have been implemented in computational models
of international politics.
International relations theory has also begun to reemphasize the importance
of nonstatistical patterns. The "international institutions" and "regimes"
literature (Krasner 1983) is quite explicit on this point. Robert Keohane, in
his 1988 presidential address to the International Studies Association, notes
institution may refer to a general pattern or categorization of activity or to a

particular human-constructed arrangement, formally or informally organized.
Examples of institutions as general patterns of behavior include Bull's "insti-
tutions of international society" as well as varied patterns of behavior as
marriage and religion, sovereign statehood, diplomacy and neutrality. . .. What
these general patterns of activity have in common with specific institutions
is that they both meet the criteria for a broad definition of institutions: both
involve persistent and connected sets of rules (formal or informal) that prescribe
behavioral roles, constrain activity, and shape expectations (Keohane 1988:
383).
While neither Keohane nor most of the international institutions literature

has provided formal definitions of these patterns, their emphasis on the
importance of pattern in international behavior is unmistakable.
This study demonstrates the use of a sequence recognition technique-
Levenshtein distances-for differentiating crises in the Behavior Correlates
of War (BCOW: Leng 1987) events data set. The importance of sequences
in determining international political behavior is discussed in general terms,
emphasizing that in a complex system such as international politics, pattern
recognition is a more efficient use of human cognitive abilities than deductive
approaches. A machine learning algorithm is developed for determining the
matrix used to discriminate the crises, and applied to the BCOW crises in
a split-sample test. An alternative formulation of the sequence recognition
problem-partially ordered event sequences-is also discussed, along with
some general comments on the differences between human and machine
understanding of politics.
Pattern Recognition and Political Behavior
Patterns Versus Axioms

Sequences of international events are useful in understanding international
politics because they can be employed as analogies. If an historical sequence
Pattern Recognition of International Event Sequences 171
of events can be found that is similar to a current sequence of events, then

that historical sequence may be useful in providing information about the
current sequence. Whether or not the analogy is actually valid is an empirical
issue: For example, in playing games, analogy is highly effective in chess
but useless in Lotto. The prevalence of the use of history in teaching
international politics (as well as in political argument) is at least prima facie
evidence that analogy is of some utility in international affairs. Decision-
makers will study the politics of the Napoleonic period for guidance on
contemporary international affairs; they do not accord the same respect to
Napoleonic surgical techniques.
lt is useful to distinguish at this point between two types of knowledge:
axiomatic knowledge and pattern recognition. Axiomatic (or deductive)
knowledge derives from first principles and is useful in understanding systems
where the relationships between variables are relatively simple and information
is relatively easy to obtain. Simple physical systems, simple economic
relationships, and artificial constructs such as mathematics, legal systems,
and bureaucracies are all susceptible to analysis by axiomatic techniques.
Because modern physical science was built on the dual foundations of
mathematics and experimentation with simple physical systems, the axiomatic
approach has been enshrined as scientific, and most attempts to systematically
study political behavior presuppose the axiomatic approach as the ideal for
systematic knowledge.
Axiomatic knowledge fails, however, when confronted with a system that
is either too complicated to be understood deductively or where critical
information is missing. This is true for physical and biological systems as
well as social systems. When axiomatic knowledge fails in a problem-solving
situation, one usually substitutes a combination of pattern recognition and
experimentation: more generally, induction. These work, ironically, because
the world is axiomatically ordered, at least at the physical and biological
levels that we experience and that are essential to survival. If one is stung
by a particular insect, one assumes it is quite likely that insects similar in
appearance will also sting. While one could derive axiomatic explanations
for this phenomenon based on complex interactions of physical, chemical,
and ecological "laws," those axiomatic explanations are not necessary as
a guide to purposeful and survival-enhancing behavior: Pattern recognition
will achieve the same results.
Whether through good fortune or evolutionary selection, the human brain
is phenomenally better at recall than reasoning. Working memory, 2 which
must be utilized in deductive reasoning, is slow and constrained to handling
only a few items of information at a time. The long-term memory used in
pattern recognition, in contrast, is effectively unlimited in capacity3 and
works very quickly-on the order of seconds-even when solving complex
associative recall problems across thousands of potential matches with an

erroneous target.
To illustrate this difference, consider the following two problems:
1. Prove the Pythagorean Theorem.

2. Who was attorney general during the 1963 Cuban Missile Crisis?
The first problem can be solved deductively using only a few simple axioms
from plane geometry, virtually every literate person has learned that proof
in high school geometry, and yet the solution is probably difficult for most
people. The second question is stated in a factually inaccurate manner,
involves a series of associative linkages and presupposes considerable
historical knowledge but can be answered quickly by most individuals
studying international politics. 4
These differences have some very specific and profound implications for
decision making about international politics. Axiomatic knowledge, while not
completely irrelevant to international behavior, is generally less important
than knowledge based in patterns and analogies. More generally, the formal
argument for using a pattern recognition approach to study international
politics can be summarized as follows:
1. International behavior results from the intentional activities of human

beings who actively seek out and process information about the
international environment.
2. Most of that activity is produced by experts who have been explicitly
and extensively trained to deal with international issues and have a
great deal of information about past international events.
3. The international system is sufficiently complex and information is
sufficiently difficult to obtain that axiomatic approaches are unlikely
to provide an adequate guide for behavior.
4. Recall is easier than reasoning in any case: A good analogy beats a
complex deductive argument in most political communications.
Therefore, regularities based on pattern recognition dominate those based

on axiomatic reasoning in determining international behavior.
Pattern Recognition and Information

Human pattern recognition of political behavior probably involves two
types of information:
1. Sequences of events;
2. The conditions at the time those events occurred.
"Conditions" are similar to conventional variables useci in statistical studies

and take on values without a temporal order; "events" are categorical and
temporally ordered. Events are activities undertaken by an actor, selected
out of a large but finite set of possible actions; conditions are environmental
variables that change slowly and affect the actions an actor is capable of
undertaking. Many conditions are quantitative, though they may be treated
as if they were qualitative, given the difficulties humans have dealing with
quantitative information (see Kahneman, Tversky, and 51ovic 1982). Conditions
are coassociated with events; they are not necessarily preconditions.
In analyzing political behavior using pattern recognition, an individual
discerns patterns by observing and classifying them as "I've seen this
before" and "I've not seen this before." Information that is seen before
reinforces existing patterns-it provides evidence that the pattern is common
and therefore worthy of attention-but is eventually discarded if it is redundant.
Information that is novel is provisionally added to memory. If it is not
reinforced, it is eventually discarded as erroneous; if it is reinforced, it is
retained. Finally, if behavior appears to match an existing pattern but further
experience shows that the match was not reliable in the sense that the
pattern predicted one outcome and a different outcome actually occurred,
then the individual either changes the matching criterion or seeks additional
information that would have distinguished the two patterns.
To take an abstract example, suppose an individual observed the following
condition-event patterns repeated on multiple occasions in a system:
5t: Cat Cbt Cct Cdt 11 Aa Ab Ac Ad Ae Af

52: Cat Cb2 Ccl Cdl 11 Aa Ab Ag Af Ae Af
53: Cat Cbt Cct Cd2 11 Aa Ab Ac Ad Af
54: Ca2 Cbt Cct Cdt 11 Aa Aa Aa Ab Ac Af
(Cxy is condition x with value y; Az is action z in the order given.) These

patterns would give the individual considerable predictive ability by using
early events in a sequence to predict later events. One could distinguish
between (5 1,53), 52, and S4 simply by looking at the first two actions (Ab
Ac, Ab Ag, or Aa Aa); to distinguish between 51 and 53 one also needs
to know the value of condition Cd. There is also a predictive symmetry
here: One can predict conditions from actions or actions from conditions.
By eliminating redundant information, these patterns could be reduced to
5t: Cat Cbt Cdt 11 Ab Ac Ad Ae Af

52: Cat Cb2 Cdt 11 Ab Ag Af Ae Af
53: Cat Cbt Cd2 11 Ab Ac Ad
54: Ca2 Cbt Cdt 11 Aa Aa Ab Ac
If the redundant information was important in some other context, it could

be reduced to a single rule that restores the information in the sequences
(e.g., "S1, . . . , S4 all have condition Cc1 and begin with Aa"). If the
patterns were commonly encountered or were associated with other very
salient patterns, they would be useful information. The long-term memory
of a human political analyst would contain thousands of such patterns.
Humans systematically learn this type of pattern recognition. Basic patterns
of human behavior are learned from infancy ("hit your sibling, you'll probably
get hit back"). Complex social patterns are initially taught through stories
and myths (and nowadays through TV); these are selectively reinforced
through social interaction as the individual matures. The complex political
patterns recognized by an expert are initially learned through the study of
a very large but highly selective set of history; these patterns are occasionally
combined into "theories" that generalize from a large number of patterns
and identify those most useful to the expert. Traumatic events provide
exceptional patterns that have a disproportionate influence on policy: For
example, in contemporary U.S. politics, decision-makers seem particularly
sensitive to event sequences that might match "Munich," "Pearl Harbor,"
or "Vietnam," just as decision-makers in the era prior to 1938 were sensitive
to sequences matching "Sarajevo."
Machine learning
Simulating pattern recognition in a computational model requires two
key components. First, one must have a nontrivial set of event sequences
in memory. Second, there must be some criteria for deciding whether two
sequences are similar. In this study I will focus only on the event sequences
themselves-without the accompanying conditions-using the Behavioral
Correlates of War (leng, 1987) international relations events data set.
Comparison will be done using Levenshtein distance (or metric), a sequence
comparison technique that originated in information theory and is now
commonly used to analyze sequences of sound or DNA.
The research approach applied to this problem comes out of the machine
learning tradition in artificial intelligence (see Forsyth and Radu 1986; Holland
et al. 1986; Michalski, Carbonell, and Mitchell 1983), and as such differs
from most of the work in this volume. The objective of machine learning
is to design systems that learn a task-in this case sequence recognition
of international behavior-rather than simulating human cognitive activity.
My earlier work on machine learning applied to sequence comparison has
included use of genetic algorithms (Schrodt 1986, 1989a), partially ordered
event sequences (Bennett and Schrodt 1987; Schrodt 1989b), and classification
trees (Schrodt 1987).
The specific problem dealt with here is discriminating between crises

that don't involve war and those that do. The Levenshtein metric uses a
large matrix of numerical weights in determining distance: The objective is
finding a set of weights that produce small distances between crisis sequences
of the same type and long distances between crisis sequences of different
types.
Training will be done using the standard machine learning protocol of
example-counterexample training. To learn to discriminate between two
classes of objects, the machine is presented with examples from each class
and adjusts its knowledge structure-in this case, the matrix of Levenshtein
weights-on the basis of those examples. Both examples and counterexamples
must be used: Most learning algorithms, presented with numerous examples
of black crows, would not conclude that "all crows are black" but that "all
objects are black" or "all objects are crows." By that same token, the class
of objects an algorithm can discriminate is dependent on the examples on
which it has been trained, the consistency of the features that account for
discrimination in the training and testing cases, and the ability of the
knowledge structure to capture the type of discrimination empirically present.
Most machine learning protocols, including this one, use split-sample
testing. Because the knowledge structures of machine learning algorithms
are very large, most systems achieve 100 percent discrimination among
their test cases. 5 They can be nontrivially tested only on data other than
those on which they were trained. In a sense, this is a difference between
learning and memorization: If a system can only parrot back the discrim-
inations in its training set, one has only demonstrated that the knowledge
structure is sufficient to "memorize" those differences, not that general
principles have been learned. In this respect machine learning studies apply
a more difficult standard of empirical accuracy than that used in most
statistical research. The accuracy of a machine learning algorithm would
be best compared to a statistical study that takes half of the data, estimates
the model, then fits the model on the remaining half of the data.
While the machine learning and statistical approaches are similar in their
use of data and empirical validation, there are also a number of differences.
First, the structures in which the information is stored are very different,
those of machine learning usually being orders of magnitude more complicated.
Second, machine learning is not based on the underlying concept of a
random variable, but rather on explicit and idiosyncratic models of the
regularities that can be expected in the data. Finally, machine learning
techniques tend to use nominal-level rather than interval-level variables,
though this is not a definitive characteristic and composite techniques exist
that use both.
There is no attempt here to achieve process validity.6 A machine is not
a human, and the best one can achieve is a method not incompatible with
human cognitive capabilities. What is gained in machine learning is the

possibility that the machine will find relationships a human researcher has
not seen.
A Levenshtein Learning Algorithm

The Levenshtein metric (Sankoff and Kruskall 1983) was first proposed
by Mefford (1984) as a means of sequence comparison of international
events. The Levenshtein distance between two sequences is the sum of the
weights of the operations needed to convert one sequence into another. If
a and b are two sequences a 1 a2 a3 • • • am and b 1 b2 b3 • • • bn, the
Levenshtein approach converts one sequence to the other using the operations
Delete an element from a

Insert an element into b
Substitute bi for a;
Using the example in Sankoff and Kruskal (1983:11), one could convert the
sequence W A T E R to W I N E by the operations
WATER
Substitute I for A
WIT ER
Substitute N for T
WINE R
Delete R
WINE
The operations used in computing the Levenshtein distance are those which
minimize the sum of the weights. A dynamic programming algorithm for
solving this minimum is presented in Figure 7.1.
The knowledge structure of a Levenshtein metric lies in the insertion,
deletion, and substitution weights. Changes in a sequence that reflect important
differences should have high weights; those reflecting trivial differences should
have low weights. For example, in linguistics, it is clear that as words
migrate from language to language, vowels are more likely to change than
consonants, and if consonants change, they are likely to change only slightly
(an "s" might change to "c" or "z" but probably not "b" or "t"). Thus
we see similarities between the English "peace," French "paix," and Latin
"pax," and between the Hebrew "shalom" and Arabic "salaam," but see
considerable differences between the two groups of words.
The extension of this principle to international event sequences is straight-
forward (Mefford 1984). Certain international events are quite comparable-
for example "mediate" (BCOW code 12142) versus "negotiate" (BCOW
Figure 7.1: Algorithm for Computing Levenshtein Distances

Function Leven dist(a,b:seq) :real;
(* a,b are arrays containing the sequences; the Oth element holds the
length of the sequence
weight[i,j) gives the insertion, deletion and substitutions
weights
dist[i,j) is the matrix used to compute the distance
This code follows the Levenshtein distance algorithm described in

Kruskal (1983) *)
var ka,t,r,c : integer;

min : real;
max r,max_c :integer;
begin
dist[O,OJ :•0.0;
for ka:-1 to a(OJ do dist[ka,O]:•dist[ka-1,0) + weight[a[ka],O);
for ka:•1 to b[OJ do dist[O,ka]:•dist[O,ka-1) + weight[O,b[ka]];
(* The code in the "t" loop goes through the matrix starting in
the upper left corner then filling by moving down and to the
left,ending at the lower right corner. r is the row, c the
column. *)
max r:•a[O);
max:=c:•b[O];
ka:-max r + max c;
for t:•2 to ka do begin
r:•l;
if t-r<max_c then c:•t-r
else begin
c:-max_c;
r:•t-c;
end;
repeat
(* Determine the operation which adds the minimum to the
weight at each point *)
if dist[r-1,c)<dist[r,c-1) then min:•dist[r-l,c)
else min:•dist[r,c-1);
if dist[r-1,c-1)<-min then min:•dist[r-1,c-1);
dist[r,c) :• min + weight[a[r],b(c));
r:•r+l;
c:•c-1;
until (c<1) or (r>max_r);
end;
Leven_dist :• dist[a[O],b(O]);
end; (* Leven_dist *I
12121)-whereas others are very different-for example "peace settlement"

(BCOW 12361) and "continuous military conflict" (BCOW 11533). One would
also expect that common events (e.g., the ubiquitous "consult" and "com-
ment" codes in events data) should be inserted and deleted with little cost,
whereas rare events such as agreements or the beginning of military conflict
would be more important. Two international event sequences would be
considered similar if one could be converted to the other with operations
that substitute like event for like event; if they could only be converted by
substituting unlike events, the two sequences would be quite different.
The problem in this approach is setting the weights. Deriving weights
on theoretical grounds would 'be difficult for a complex coding scheme: For
example, what should be the relationship between BCOW 13551 ("reach
economic agreement") and BCOW 12641 ("assume foreign kingship")? The
alternative is induction: weights determined by what one wants to do with
the Levenshtein measure itself.
The algorithm I've developed for determining weights is based on the
Widrow-Hoff or "delta rule" training method used in training neural networks. 7
The machine is given cases in two categories: war crises and nonwar crises
from the BCOW set. The training objective is creating weights that will
produce small distances between the sequences within each set, and larger
distances between sequences in different sets.
To create the weights, the distances between each pair of sequences is
computed using the Levenshtein algorithm. Any weights used in computing
the distance between a pair of like sequences are decreased by a small
amount. Weights used in computing the distance between unlike sequences
are increased by the same amount. The weights of operations invoked in
comparing both like and unlike sequences remain the same, since the
increase and decrease cancel out. As a consequence, the distances within
the groups should decrease, while the distances between the groups should
increase. The learning has to be done iteratively since the choice of operations
used in computing the Levenshtein distance may change as the weights
change.
In the experiments described below, the weights were initialized in two
different ways. Frequency-based initializations set the insertion and deletion
weights to the rank-order of the event frequency in the set of all sequences
used to train the system. The most frequent event had a weight of 1, the
second most frequent a weight of 2, and so forth. This is consistent with
the coding used in my earlier work and is based on the information theoretic
argument (Pierce 1980) that frequent events have little discriminating value
and should be replaced with little cost, whereas rare events should have a
higher cost. 8 The substitution cost was initialized as lr. -rbl, the absolute
difference of the ranks of codes a and b. Thus it is less costly to replace
a frequent event with another frequent event than it is to replace a frequent
event with a less frequent event. Alternatively, weights were initialized to a
constant. These different initializations had no effect on the learning algorithm,
though the constant weights proved somewhat less useful for doing dis-
crimination.
The learning scheme produces a few peculiarities in terms of the usual
mathematical regularities expected of a "distance" (Sankoff and Kruskall
1983: 22), but these cause no interpretive problems in the discrimination
test. First, the metric is completely arbitrary, without a zero point, and the
"distance" may be negative since many weights become less than zero by
progressive subtraction when their operations are repeatedly used in com-
paring like sequences. For that same reason, the distance between two
identical sequences is not necessarily zero: In fact it tends to be negative.
Finally, unless the weight matrix is symmetric, the distance between A
and B is not necessarily the same as the distance between B and A. Because
the changes in weights are done while the sequences are compared, a
different weight matrix is used to compare A to B than comparing B to A
during the training. As a consequence, the weight matrix is not symmetric.
There are no substantive excuses for this: It is simply a quirk in the
algorithm.
Testing the Algorithm
Data
This system was tested using the BCOW sequences studied in Schrodt
(1989b). BCOW focuses on a limited number of historical crises using a
variety of historical sources, so it has a greater number of events per unit
time than WEIS or COPDAB, which attempt to cover the entire world using
a single information source. While BCOW codes both physical and verbal
activity, I analyzed only the physical actions on the assumption that these
would be more regular over time and across cultures than verbal actions.
The focus on physical actions considerably shorten the sequences: This
was important due to memory constraints and because time required to
compute a Levenshtein distance is proportional to the product of the length
of the two sequences. The four subsets of crises listed in Table 7.1 were
analyzed. 9 The short names (e.g., "pastry") correspond to the BCOW file
identifiers. "Training" sequences were used to teach the system to discriminate
between war and nonwar sequences; the system was tested with the remaining
sequences. 10
The sequences were filtered on the basis of novelty to eliminate common
events such as consultations, accusations, and, in the case of wars, acts
of violence. These high-frequency events are noise in the sense that they
can be eliminated without loss of information about the sequence because
they are occurring with a much higher frequency than the frequency of
the events that are important in determining the classification of the sequence.
The ongoing shouting matches and violence in BCOW mask the slower,
more significant processes of military escalation or diplomatic rapprochement
and are the events data equivalent of the static from a lightning storm
intruding on a radio broadcast of the "Toccata and Fugue in D."
Table 7.1 : Data Sets Analyzed
Crise5 without war
Training Set
BCOW file Crisis Date Length.il

fashod Fashoda Crisis 1898-1899 32
1stmor First Moroccan Crisis 1904-1906 79
bosnia Bosnian Crisis 1908-1909 116
2ndmor Second Moroccan Crisis (Agadir) 1911 38
rhine Rhineland Crisis 1936 65
Test Set
pastry Pastry War Crisis 1838-1839 41
brprt British-Portuguese Crisis 1889-1890 15
anschl Anschluss Crisis 1937-1938 37
munich Munich Crisis 1938 114
berair Berlin Blockade 1948-1949 118
Crises inyolyiog war
Training Set
BCOW file Crisis Date Length.il
schles Schleswig-Holstein War 1863-1864 52

spam Spanish-American War 1897-1898 171
cent am Second Central American War 1906-1907 71
chaco Chace Dispute and War 1927-1930 125
italet Italo-Ethiopian War 1935-1936 260
Iest Set
balkan Balkan Wars 1912-1913 115
palest Palestine War 1947-1948 177

kashl First Kashmir War 1947-1949 70
kash2 Second Kashmir War 1964-1966 76
bang la Bangladesh War 1971 108
aLength number of events in the filtered sequence
Discriminating War and Nonwar Crises

The basic protocol was to run the algorithm on the ten training cases,
iterating until the two groups were separated. 11 The resulting weight matrix
was then applied to the ten test cases by computing the distance between
each test case and the ten training cases. The expectation was that the
nonwar cases would, on average, be closer to the nonwar cases in the
Table 7.2: Distances in Training Set
rh in 1mor fash 2mor bosn sch1 spam cent ital chac

rhine: 330 144 139 311 603 1771 1081 1946 1446
1stmor 190. -148 189 -275 532 1540 1162 1577 1393
fa shod -288 -77 -177 308 437 1496 754 690 803
2ndm 203 142 -168 351 479 1154 628 1298 972
bosnia 258 -155 307 337 773 -467 662 1493 521
schles 358 432 561 389 540 -2874 -2286 -1301 -1080
spam 1413 1432 1521 1407 -533 -3277 -6399 -5747 -6724
cent am 1064 1158 773 613 421 -2371 -6277 -3363 -4641
ita1et 1903 1781 1586 1648 1594 -1357 -4941 -3145 -3216
chaco: 1184 1154 1028 715 474 -1346 -5076 -3161 -2510
Average within-group distance

Average between-group distance
- -1729.4
987.6
Separation 2717.0
Table 7.3: Distances in Test Sets
~~1: di~tiDC~ ~DU~ DQDJIIII: di:ltiDCS:

Cz:i:ai:s I¥ge fte~eccx CQD:ItiDt
palest War -2253 -4494
balkan War -1264 -4189
bang la War -994 -2932
kash1 War -624 -3685
kash2 War -557 -3226
munich Nonwar 21 -708
berair Nonwar 210 291
pastry Nonwar 436 -821
anschl Non war 713 316
brtprt Nonwar 1645 552
training set, and similarly for the war cases. The training algorithm showed
a monotonic increase in the separation of two groups. This increase in
separation was mostly linear with a very slight leveling-off that one would
expect in a classical learning curve.
Table 7.2 gives the results of testing the comparisons with the training
cases. 12 The expectation that the Levenshtein distance would discriminate
between the war and nonwar crises is fully met, and the discrimination is
almost perfect. Only one crisis-Schleswig-H olstein-is not strongly classified
into the correct group. The war crises cluster strongly, with large negative
distances within the group and large positive distances between the groups.
Table 7.3 reports the split-sample test using the difference in distances:
(Average distance to war crises) minus (Average distance to nonwar crises)
Results for both frequency-based and constant initializations are reported.

As Table 7.3 indicates, the discrimination using the frequency-based ini-
tialization is perfect: All of the war crises in the target set are closer to
the war crises in the training set than to the nonwar crises in the training
set; the reverse is true for the nonwar crises. In terms of the rank-order
of the distances, the same is true when constant initial weights are used
but two of the nonwar crises are actually closer to the war group than to
the nonwar group. In this test, frequency-based initialization seems to produce
a better discrimination matrix than constant initialization.
Table 7.4 shows the event pairs that had the maximum and minimum
changes in substitution weights from their initial values. 13 Neither group is
particularly surprising. The minimum weights, which indicate strong similarity,
are usually similar or identical events, particularly those involving military
conflict. The maximum weights, which indicate strong dissimilarity, primarily
involve substitution of a cooperative action such as consultation or negotiation
for a military action such as "show of strength" or "take POWs." The
magnitude of the largest minimum weights is substantially greater-almost
by a factor of ten-than the magnitude of the maximum weights. The first
digit is a dyad identification code so, for example, the "mobilization/
mobilization" pair in the maximum weights is the substitution of a mobilization
by an "other" against one of the principals for a mobilization by one of
the principals against the other. (Jnsurprisingly, most of the large weights
occur in the frequent events, and there were thousands of substitutions that
never changed from their initial value. 14
Table 7.5 reports the insertion and deletion weights for the most frequent
events. These tend to be symmetric for insertion and deletion, inversely
proportional to their rank orderings and substantially higher than the
substitution weights. This pattern holds whether frequency-based or constant
initial weights are used and is opposite the information theoretic that frequent
events would have the smallest insertion and deletion weights. These high
values imply that at some point in the training process, insertions and
deletions were being used frequently-otherwise they would not have the
high values-but their high values relative to the substitution weights would
lead one to expect that eventually the sequence comparisons would be
dominated by substitutions. A few codes showed negative weights, usually
on the order of 10 to 100, but almost all were positive.
The Levenshtein learning algorithm is clearly only a first step in the
larger puzzle of learning to deal with international events as sequences.
Alone, it recalls Faraday's retort to the lady who enquired about the utility
of his experiments with electricity-"What is the use of a newborn baby?"-
albeit I've no illusions that this algorithm is comparable in utility to electricity.
Table 7.4: Minimum and Maximum Weights
Minimum Weights
Code Meaning Code Meaning weight

111633 Military victory 111633 Military victory -687.5
111513 Clash 111633 Military victory -579.6
121143 Change in force 121143 Change in force -446.6
111663 Take POWs 111523 Attack -299.1
111533 Continuous conflict 111533 Continuous conflict -278.3
111523 Attack 111523 Attack -237.6
211131 Military coordin. 112111 Consult -162.7
312521 Reach Agreement 411313 Show of Strength -128.6
111553 No code 512111 Consult -119.8
112111 Consult 112111 Consult -111.1
312213 Violate Territory 111313 Show of Strength -109.9
212521 Reach Agreement 412111 Consult -103.3
211553 No code 111353 Mobilization -103.3
312142 Mediate 111533 Continuous conflict -94.5
112631 Attend Internatnl Event 123151 Change Trade -90.1
Maximum Weights
Code Meaning Code Meaning Weight

512111 Consult 111313 Show of Strength 74.9
212521 Reach Agreement 111663 Take POWs 70.5
112121 Negotiate 512111 Consult 69.4
111663 Take POWs 112121 Negotiate 69.4
114113 Subversion 111313 Show of Strength 69.4
112213 Violate Internatnl Law 111313 Show of Strength 68.3
212111 Consult 111633 Military victory 67.2
111353 Mobilization 311353 Mobilization 67.2
312173 Expel Foreign Rep 112213 Violate Internatnl Law66.1
321133 Change Force Level 512111 Consult 63.9
311313 Show of Strength 512111 Consult 60.6
112521 Reach Agreement 514143 Assassinate 60.6
111443 Military Intrusion 111663 Take POWs 59.5
111523 Attack 112111 Consult 58.4
111513 Clash 112111 Consult 57.3
The strength of the approach lies in the inductive nature. There are clearly
simpler rules for distinguishing BCOW war and nonwar crises: Looking for
codes involving military conflict is the most obvious. But in order to construct
those simpler rules, one must first know the distinguishing characteristic
one is looking for: In a sense, one must already know the answer. An
inductive learning algorithm does not need to know the answer: lt will find
the answer. The system did not know, a priori, the importance of the BCOW
codes designating military conflict; it discovered them. If machine learning
Table 7.5: Insertion and Deletion Weights
biszbta
Cod,e Ht:aniog Delete IDiiU::t
212111 Consult 2608.1 2638.9
312111 Consult 239)~ 6 2545.4
512111 Consult 1548.8 1951.4
111313 Show of Strenqth 2208.8 2077.9
111353 Mobilization 1763.3 2124.1
111633 Military victory 2083.4 1675.3
112111 Consult 862.4 652.3
111523 Attack 1180.3 1020.8
311313 Show of Strenqth 597.3 108.8
112121 Neqotiate 731.5 735.9
112521 Reach Aqreement 994.4 906.4
111653 Occupation 895.4 914.1
111513 Clash 1298.0 1285.9
212121 Neqotiate 789.8 699.6
111663 Take POWs 902.0 412.5
systems can discover those distinctions, they may be capable of discovering

things that are not so obvious. 15
Partially Ordered Event Sequences

There are two major weakness of the Levenshtein approach. First, it
can't deal with permutations. Second, it only measures distances between
known sequences: There is no way to present, in the abstract, an archetypical
sequence characteristic of a class of sequences. One approach to this more
general problem that does not use Levenshtein distances follows work by
Heise (1988a, 1988b; Coraso and Heise 1987), who has developed it for the
study of general human interactions, for example, ethnographic data, stories,
or fairy tales. Computational techniques for applying the approach using
machine induction are developed in Bennett and Schrodt (1987) and Schrodt
(1989); Heise (1988b) has been applying the method with machine-assisted
techniques.
Heise's partially ordered event structures (POES) provide sequences in
which for certain events to occur, other events must occur first to prime
the system for the consequent event. However, because an event may have
multiple preconditions, and the preconditions can be satisfied in any order
so long as they occur before the consequent event, the sequence is only
partially ordered.
A partially ordered structure is illustrated in Figure 7.2. This would
generate any of the following "Breakfast" sequences:
Figure 7.2: A Partially Ordered Event Structure
Drink Coffee
I
Pour Cereal
I Cook
I
Toast
Coof Eggs
I I
Butter Toast
Eat
I
Wash Dishes
[Drink Coffee] - (Cook Eggs] - [Pour Cereal] - [Cook Toast] - [Butter Toast]
- [Eat] - [Wash Dishes]
or
[Drink Coffee] - [Cook Toast] - [Butter Toast] - [Pour Cereal] - [Cook Eggs]
- (Eat] - [Wash Dishes)
or
[Drink Coffee] - [Cook Toast] - [Cook Eggs] - [Butter Toast] - [Pour Cereal)
- [Eat) - [Wash Dishes]
but would not generate
[Cook Eggs] - [Pour Cereal) - [Cook Toast] - [Butter Toast] - [Drink Coffee]
- (Eat] - [Wash Dishes]
or
[Drink Coffee] - [Butter Toast] - [Pour Cereal] - [Cook Toast] - [Cook Eggs]
- [Eat] - [Wash Dishes]
because in each of the latter examples one of the events occurs out of the
allowed sequence.
POES seem a more accurate representation of human intentional activity
than the linearly ordered sequences that can be compared using Levenshtein
metrics. Humans have objectives that require an oftentimes complex set of
preconditions. In most human situations at least some of the preconditions
are partially ordered, and the temporal ordering of the occurrence of specific
events in a given partial order is often very random.
The extension of POES to international affairs is straightforward: When

states seek to achieve an objective, they initiate an often complex sequence
of events that will result, ceteris paribus, in the objective. For example, if
a state wishes to initiate peace talks, it will contact its allies, the United
Nations, initiate trial balloon proposals, and so forth. In international relations,
these initiatives are often not fully completed because they are interrupted
by opposing initiatives by other states or through changes in policy, but
the plan was there. Such interruptions also provide a reason to expect that
there will be greater short-term regularity than long-term regularity in
international events, since many event sequences are attempted and fail for
every sequence that succeeds.
POES provide a form of event grammar. The partially ordered structure
is a grammar in the sense that it can be used to generate a set of well-
formed sequences. The concept of an event grammar is by now fairly
common: A useful survey of various "story grammar" concepts is found
in Alker (1987). The initial work in Bennett and Schrodt (1987) and Schrodt
(1989) found that POES models could successfully match an almost em-
barrassing percentage of events data in WEtS and BCOW. In both studies,
about ten POES could account for about 35 percent of the data. In Bennett
and Schrodt (1987), the sequences were rather boring-mostly repetitive
strings of violence, comments, and accusations-but in the more detailed
BCOW set studied in Schrodt (1989b), some of the POES captured the
common outlines of various phases of crisis behavior (mobilization, nego-
tiation, and so forth) The BCOW sequences showed some ability to dis-
criminate between war and nonwar crises both in terms of informal inspection
of the sequences and also in a formal test using correspondence analysis.
Unfortunately, the computational methods I am using with POES do not
deal easily with event substitution-the forte of Levenshtein metrics-so
for the moment the two techniques are complementary.
The construction of archetypical or "ideal" sequences is another interesting
problem for which I've yet to find a computational solution. In human pattern
recognition, we have a general idea of what category event sequences look
like-the archetypical war, the archetypical coup, and so forth-and probably
match to these ideals rather than to clusters of sequences. In a sense, ideal
sequences are the centroid of a cluster of sequences, but the centroid is
a sequence rather than a point. If a means could be found for constructing
such a sequence, the cluster could be represented by the single ideal
sequence, which would substantially reduce computing time and provide
some theoretical insights as to the distinguishing characteristics of a cluster.
But Is It AI?
Outside the machine learning literature, where empirical testing is the
norm, computational techniques that lack even the pretension of process
validity and that use actual data are often considered merely statistical and
alien to the realm of artificial intelligence. While the boundaries of AI may,
in the final analysis, seem like an issue of theology rather than methodology,
a few words are in order on this subject. In particular, this question raises
some interesting issues as to what AI should expect when applied to the
study of political behavior.
I contend this method is a form of Al. First and foremost, it uses the
iterative example-counterexample protocol of machine learning: It learns.
Second, it uses a large and highly diffuse knowledge structure combined
with a comparison technique specified as an algorithm rather than as an
equation; neither is characteristic of statistics. Finally, and least important,
the algorithm is based on a neural network technique, the Widrow-Hoff
algorithm. To the extent that machine learning, knowledge-intensive algo-
rithmic problem solving and neural networks are AI, the model is based in
that literature.
The algorithm is also based on an explicit model of human reasoning
about international affairs. Most of the psychological evidence indicates that
the human agents producing international behavior are highly sophisticated
at pattern recognition: This project seeks to apply pattern recognition in a
world of behaviors produced by pattern recognizers. Event sequences appear
to be important in human reasoning about international politics; to the best
of my knowledge, this is the only technique in IR that has directly compared
entire sequences of events data without aggregating them. The method
discussed here is specific to international events and would be utterly useless
in predicting presidential elections, approving credit limits, or diagnosing
heart disease.
I would argue that the term "artificial intelligence" as applied to the study
of international affairs is a misnomer if "intelligence" means "human
intelligence." It is useful if "intelligence" means complex information pro-
cessing of which human intelligence is one example, but not a unique
example. If one can speak of "insect intelligence" then one can speak of
"machine intelligence." Just as one trains a child to catch a Frisbee differently
than one trains a dog to catch a Frisbee, so it may be necessary to teach
a machine to "think" about politics differently than one would teach a
human.
While both humans and machines can exhibit intelligence, the contrast
between the mechanisms of human and machine intelligence is striking and
must be taken into consideration when constructing computational theories
of politics. Humans are poor at the serial, logical processing of information
and exceptionally good at associative recall; machines are exceptionally
good at logical processing but poor at associative recall. Machines are at
a further disadvantage in having very little background information about
patterns of human behavior, whereas humans accumulate this information
from infancy. Consequently, a theory that makes sense for a human being
will not necessarily make sense to a machine. The term "makes sense" is
here defined as a theory that can be transmitted and understood by another
and used to predict or guide behavior.
Comprehension and utilization of political theory are a combination of
both transmission and experience. 16 lt is not sufficient simply to convey
information; instead that information must link with other information that
is common between the sender and receiver before it can be used. As
twentieth-century humans, we find our political environment has enough in
common with the environments of Thucydides' Greece, Kautilya's India, and
Sun-Tzu's China-to say nothing of the Europe of Marx and Weber-that
we can understand their theories. 17
Computational models of politics need to be consistent with human
cognitive abilities but, in order to deal with real problems, must also be
able to be processed using a machine. The point is not to make the machine
into a human being, but rather to make the machine solve problems similar
to those solved by human beings, but in its own way. As such, theories of
political behavior designed to be understood by humans with access to a
range of common human experience are of little use to machines in their
raw form, though they probably have significant heuristic value. Specialized
theories that are designed for machines are needed. Duplicating human
comprehension is impossible for the foreseeable future, up until such time
we provide machines with the full gamut of human experience, including
adolescence (particularly adolescence . . . ), but predicting human behavior
is quite possible, just as we can predict the behavior of chickens without
being chickens.
The field of rule-based expert systems provides two useful illustrations
for the contrast of human intelligence and machine intelligence. Expert
systems are easily capable of replacing human intelligence (at an equal or
superior level of performance) in two general areas: classification problems
(e.g., diagnosis of disease, repair, and configuration of equipment) and
duplicating the information processing of large bureaucracies (e.g., insurance
underwriting, the American Express credit approval system). In the case of
classification problems, the computer is clearly not solving the problem in

a human fashion. The information required for an expert problem-solving
system is reduced to rules by human "information engineers" or using
machine learning techniques; the simplicity of the resulting systems are
usually of considerable (and often demoralizing) surprise to the human expert.
The human expert works with associative recall over a large, unsystematic
but representative set of examples rather than through the logical processing
of a relatively small number of rules. Chess-playing machines and humans
also function at comparable levels of expertise (except for the decreasing
number of humans who can still play chess better than a machine), but
almost no one now argues that machine chess-playing duplicates the cognitive
processes of human chess-playing (Simon 1982: 107). These machine
intelligence systems duplicate the performance of a human expert without
duplicating the process.
Machine replacement of bureaucracy, in contrast, largely duplicates the
formal, rule-based structure of bureaucratic decision making in a machine.
It duplicates the process as well as the performance. This is possible because
a bureaucracy, unlike an individual human expert, has an information-
processing structure similar to that of a computer. Bureaucracies achieve
a high level of processing by coordinating in parallel the activity of a large
number of humans with low logical processing ability. A bureaucracy also
has a relatively low amount of associative memory: Because of high turnover
and the expense of training, most organizational memory is in physical
information files, which are generally not associatively accessed. A bureau-
cracy, in short manages to reduce human information processing to the
level of a computer, hence a computer can duplicate the process. For that
subset of political activity that is the product of bureaucracy, one can
therefore construct a model with some process validity: the JESSE simulation
(Sylvan, Gael, and Chandrasekaran 1988, this volume) is the obvious example;
Job, Johnson, and Selbin's (1987, this volume) simulation of U.S. Central
American policy is another.
Human information processing dealing with international affairs clearly
involves more than just event sequences. However, event sequences are an
important element of that processing, and, just as importantly, sequences
are sufficiently concrete that machines can deal with them. In addition, a
very large amount of information on international event sequences is already
available in machine-readable form. These data have a number of well-
known problems (ISQ 1983) but they are probably still sufficiently close to
reality that they can be used for theoretical and methodological development;
those theories and methods could then motivate better events data collection
efforts. While events data have not fared particularly well when subjected
to statistical analysis, they may prove considerably more useful when analyzed
in a manner more comparable to how humans view events.
Notes
1. The bulk of the AI literature is concerned with spatial pattern recognition-
recognizing a flaw in a gear or distinguishing a tank from a truck-but develops a
number of general concepts and techniques that carry over to the temporal patterns
of international behavior.
2. Earlier known as "short-term memory."
3. See Newell and Simon (1972), chapter 14. Newell and Simon argue that the
capacity of associative memory is effectively unlimited because the amount of time
required to store items is sufficiently long that life span, rather than memory capacity,
is the constraint. What is stored in that memory is a separate issue, of course, and
obviously a decision-maker will carry a lot of information in associative memory
that is irrelevant to most international problems: e.g., the current price of bananas,
the last episode of Dallas, the name of her date to the high school senior prom,
and so forth.
4. Robert Kennedy, and the crisis was in 1962, not 1963. The fact that the
attorney general was the president's brother and actively involved in the crisis aids
in the recall; I suspect most people could not answer the same question for the U-
2 crisis.
5. The exception occurs when two cases have different classifications but have
identical values for all of the independent variables. In such situations insufficient
information exists in the data set to make the discrimination.
6. As discussed in the final section, achieving process validity in human learning
about political behavior is virtually impossible because such learning, in humans,
presupposes knowledge of human social behavior, and much of that is learned
informally (e.g., through family interactions). It would be an interesting experiment
to have a machine learn the event sequences found in the primary school history
textbooks used in, for example, the United States, Nigeria, the Soviet Union, and
Japan and see whether significant differences emerged. Such a project is impractically
expensive, but I would guess that at this elementary level, where the most basic
formal political sequences are first learned, one would find significant cross-cultural
divergence.
7. The original method was discussed in Widrow and Hoff (1960); Rumelhart et
al. (1986) provide an extensive discussion of variations in the context of neural
networks.
8. No adjustment was made for ties: Events tied in frequency were randomly
ordered within that tie. BCOW events generally follow a rank-size law in their
distribution.
9. The BCOW crises not included in the study are generally those whose length
in events is very long (e.g., Suez or the Cuban Missile Crisis) or those I could not
easily classify into war or nonwar (e.g., Trieste). No deliberate attempt was made
to manipulate the results by choice of crises, though I did deliberately choose the
training cases to be representative of the test cases.
10. Letting "principals" refer to the Side A and Side B designated in BCOW,
any other state as "other," the prefixes are: 1 =interaction between principals;
2=principal as initiator, other as target; 3=other as initiator, principal as target;
4=interaction within principals (e.g., between actors on same side); 5=interaction
between others. These codes were prefixed to the BCOW code: for example 112111
is "diplomatic consultation between Side A and Side B" and 312111 is "diplomatic
consultation initiated by an 'other' and directed to either Side A or Side B."
11. The algorithm was implemented in Turbo Pascal and run on a Macintosh ll
computer. The programs are available "as is" from the author: They are ordinary
Pascal and should run with little conversion on MS-DOS machines, though they use
considerable memory.
12. Table 7.2 was generated with 20 iterations of the training set, constant initial
weights set at 10.0, and the weight increment of 1.1.
13. The weights in Tables 7.4 and 7.5 were produced with 51 iterations under
the same conditions as Table 7.2. The reported weights were selected from the set
of maximum and minimum substitution weights for each event rather than from
the maximum and minimum events for the entire table.
14. There were 171 distinct dyad-prefixed event codes in the sequences, so the
total size of the matrix, including the insertion and deletion weights, was 1722, or
29,584.
15. I have also experimented with the discrimination of multiple categories. The
results are encouraging, though not spectacular.
16. A related and very thorny issue in the AI community is whether machines
have the ability to utilize natural language to a significant degree. My views on this
are presented in detail in Schrodt (1988): In a word, I am very skeptical.
17. An extreme cultural relativist would argue that true "understanding" is never
achieved between individuals of different cultures because of these differences. In
the strictest sense, this is true, but the issue is one of degree, not kind. Environments
are never identical between individuals, much less societies: J. Danforth Quayle and
I both grew up in Indiana, in the United States, in a late twentieth-century Anglo-
American political culture, and yet many of our fundamental assumptions about
political culture are doubtlessly (hopefully ... ) profoundly different. In contrast, one
can go to an area with a totally different political culture and language and with
time and effort still /earn to internalize enough of that environment to be able to
"understand" that culture sufficiently well to make intelligent statements about it.
The validity of these statements can be ascertained either culturally by the agreement
of individuals who are native to that culture (the same criterion used in ascertaining
whether a sentence is grammatically correct: agreement by a native speaker) or
empirically by making successful predictions.
I've noticed that when one becomes involved in serious firsthand cross-cultural
political exchange (as opposed to debate), there is a tremendous emphasis on
information that can be used to construct an accurate model of how the other
political system works ("No, no, that's not what we would do, and here's why
. . ."). The process is eerily similar to a machine learning protocol. While such
communicated understanding is never perfect, much can be learned, and it is greatly
facilitated by the accumulation of information and experience within the culture:
Nothing beats being there. One of the most ignorant assumptions of the 1960s
behavioralists was that this process of area-specific knowledge acquisition could be
bypassed: This reflected the assumption that political understanding could bypass
the use of associative memory and be reduced to a few axioms that would be
processed serially.
Bibliography
Alker, Hayward R. 1987. "Fairy Tales, Tragedies and World Histories: Towards
Interpretive Story Grammars as Possibilist World Models." Behaviormetrika 21:
1-28.
Bennett, Scott and Philip A. Schrodt. 1987. "Linear Event Patterns in WEIS Data."
American Political Science Association, Chicago.
Corsaro, William A. and David R. Heise. 1987. "Event Structure Models from
Ethnographic Data." Mimeo, Department of Sociology, Indiana University.
Devijver, Pierre A. and Josef Kittler (eds.) 1987. Pattern Recognition Theory and
Applications. Berlin: Springer Verlag
Forsyth, Richard and Roy Rada. 1986. Machine Learning: Applications in Expert
Systems and Information Retrieval. New York: Wiley fHalstead.
Fu, K.S. 1974. Syntactic Methods in Pattern Recognition. New York: Academic Press.
___ . 1982. Syntactic Pattern Recognition and Applications. NY: Prentice-Hall.
Gochman, Charles S. and Russell J. Leng. 1983. "Realpolitik and the Road to War:
An Analysis of Attributes and Behavior." International Studies Quarterly 27,1:
97-120.
Heise, David. 1988a. "Modeling Event Structures." Journal of Mathematical Sociology
13: 138-168.
___ . 1988b."Computer Analysis of Cultural Structures." Social Science Computer
Review 6: 183-196.
Holland, John H., Keith J. Holyoak, Richard E. Nisbett, and Paul R. Thagard. 1986.
Induction: Processes of Inference, Learning and Discovery. Cambridge: MIT Press.
International Studies Quarterly. 1983. "Symposium: Events Data Collection." Inter-
national Studies Quarterly 27.
Job, Brian L, Douglas Johnson, and Eric Selbin. 1987. "A Multi-Agent, Script-based
Model of U.S. Foreign Policy Towards Central America." American Political Science
Association, Chicago.
Kahneman, Daniel, Paul Slovic, and Amos Tversky. 1982. Judgement Under Uncer-
tainty: Heuristics and Biases. Cambridge: Cambridge University Press.
Keohane, Robert. 1988. "International Institutions: Two Approaches." International
Studies Quarterly 32: 379-396.
Kohonen, T. 1984. Self-Organization and Associative Memory. New York: Springer
Verlag.
Krasner, Stephen D. (ed.) 1983. International Regimes. lthaca: Cornell University
Press.
Kruskal, Joseph B. 1983. "An Overview of Sequence Comparison," pp. 1-44 in
Sankoff, David and Joseph B. Kruskal (eds.), Time Warps, String Edits and
Macromolecules: The Theory and Practice of Sequence Comparison. New York:

Addison-Wesley.
Leng, Russell J. 1987. "Behavioral Correlates of War, 1816-1975." (ICPSR 8606).
Ann Arbor: Inter-University Consortium for Political and Social Research.
Margolis, Howard. 1987. Patterns, Thinking and Cognition: A Theory of Judgement.
Chicago: University of Chicago Press.
McCielland, Charles A. 1976. World Event/Interaction Survey Codebook. (ICPSR
5211). Ann Arbor: Inter-University Consortium for Political and Social Research.
Mefford, Dwain. 1984. "Formulating Foreign Policy on the Basis of Historical Analogies:
An Application of Developments in Artificial Intelligence." International Studies
Association, Atlanta.
Michalski, Ryszaid S., Jaime G. Carbonell, and Tom M. Mitchell. 1983. Machine
Learning: An Artificial Intelligence Approach. Palo Alto: Tioga Publishing.
Newell, Alien and Herbert Simon. 1972. Human Problem Solving. Englewood Cliffs,
NJ: Prentice-Hall.
Patrick, Edward A. and James M. Fattu. 1986. Artificial Intelligence with Statistical
Pattern Recognition. Englewood Cliffs, NJ: Prentice-Hall.
Pierce, John R. 1980. An Introduction to Information Theory (2nd edition). New
York: Dover Publications.
Rumelhart, David E., James L. McCielland, and the POP Research Group. 1986.
Parallel Distributed Processing (2 vols.). Cambridge: MIT Press.
Sankoff, David and Joseph B. Kruskal (eds.). 1983. Time Warps, String Edits and
Macromolecules: The Theory and Practice of Sequence Comparison. New York:
Addison-Wesley.
Schrodt, Philip A. 1984. "Artificial Intelligence and International Crisis: An Application
of Pattern Recognition." American Political Science Association, Washington, DC.
___ . 1985. Precedent-Based Logic and Rational Choice: A Comparison, in Michael
Don Ward and Urs Luterbacher (eds.), Dynamic Models of International Conflict.
Boulder, CO: Lynne Rienner.
___ . 1986. "Predicting International Events." Byte 11,12 (November, 1986)
___ . 1987. "Classification of Interstate Conflict Outcomes using a Bootstrapped
CLS Algorithm." International Studies Association, Washington, DC, April 1987.
___ . 1988. "Language, Pattern and Computational Models of Political Behavior."
Shambaugh Conference, University of Iowa.
___ . 1989a. "Short Term Prediction of International Events using a Holland
Classifier." Mathematical and Computer Modelling 12: 589-600.
- - - . 1989b. "Parallel Event Sequences in International Crises." Political Behavior.
Forthcoming.
Simon, Herbert A. 1982. The Sciences of the Artificial (2nd edition). Cambridge:
MIT Press.
Sylvan, Donald A., Ashok Goel, and B. Chandrasekaran. 1988. "An Information
Processing Model of Japanese Foreign and Energy Policy Decision Making: JESSE."
International Studies Association, St. Louis.
Widrow, G. and Hoff, M.E. 1960. "Adaptive Switching Circuits." Institute of Radio
Engineers, Western Electronic Show and Convention, Convention Record, Part 4:
96-104.
8
Scripting International Power Dramas:
A Model of Situational Predisposition
Valerie M. Hudson
These days it is difficult, though not impossible, to assert that nation-

states behave in a classically rational fashion. Numerous variants of the
classic formulation have been offered by students of international relations
to account for discrepancies in observed nation-state behavior. Bueno de
Mesquita (1981), for instance, offers the risk orientation of the national
leader as a variable important in determining decision rules for expected
utility maximization; Achen (1989) offers a reformulation of the unitary
rational actor model that incorporates insights from bureaucratic politics
literature; McGinnis and Williams (1989) redefine the decision-maker's task
as juggling a two-level game (domestic and international).
Others, such as Etheredge (1985, 1987), have argued that what is needed
is not a reworking of rational actor models but a supplementation of those
models. He puts forth a framework that would treat decision-makers as
"dual track information processors." The first track would correspond to
common models of bounded rationality. The second track would comprise
the mind's imaginative coding of events in dramatic form. This second
track suggests that international relations are not only a more or less rational
game played out between actors but that it is also high drama, with actors
constrained by the dramatic requirements of the role they play. The give-
and-take of international relations becomes a series of impression management
policies, within which the trace of rational calculus can be found. Dramatic
requirements first set the form the policy must take; rationality then defines
the precise substance that will occupy that form. Etheredge argues it is
the precedence of the former over the latter that is often unrecognized in
rational choice literature in international relations, leading researchers to
undertake strenuous conceptual contortions in their attempts to fit observed
nation-state behavior into the rational actor model.
Let's assume for the body of this chapter that Etheredge is correct-
that to understand and project nation-state behavior we must first know
194
Scripting International Power Dramas 195
the script of the drama they are enacting. This idea is not new to those
in the AI community (see Schank and Abelson, 1977). Is it possible to
develop a conceptual framework that would tell what the current "scene"
is and what the dramatic requirements for a particular nation playing a
particular role in the scene are? lt would then be a straightforward task to
see if the nation in fact behaves as its role requires. If successful, this
modeJI would provide the situational predisposition of nations at discrete
temporal points; that is, it would estimate how the power drama predisposes
an acting nation (the actor) in any specific situation (the scene) to behave
(the foreign policy behavior). I believe it is quite possible to construct such
a framework. Indeed, I feel our commonsense understanding of international
politics, the explanations of foreign policy given by government officials,
and the analyses offered by journalists all point to the fact that most of
us are capable of "scripting" international power dramas. The challenge is
to do so in a systematic fashion, allowing for cross-national application and
rigorous empirical testing. This chapter posits such a framework and offers
an empirical test for twenty-eight nations over the span of a decade.
The Conceptual Framework

Given that our natural metaphor is the theater, it is not surprising that
the theoretical wellsprings of the conceptual framework (hereafter called the
situational predisposition model) should be found in role theory.
Role theory is an approach derived from the theater; it has been used
in various forms by anthropologists, sociologists, and psychologists. The
basic premise of role theory is that social behavior is in large part a function
of the expectations attached to or associated with individuals on the basis
of their locations or positions in social systems. Around this position converges
a collection of rights and duties. A role, then, is a status put into action.
Taking from the work of such scholars as Linton (1945), Gross (1958),
Nadel (1957), Banton (1965), Goodenough (1965), and Goffman (1961),
behavior of individuals in a social system can be explained by reference to
the concept of "role behavior"-i.e., actual performance that relates to the
norms and expectations of the role of the individual. This role behavior is
enacted within the context of a "role set," that is, the set of relevant others
and their roles within the system. In practice, the role expectations are in
fact often the researcher's, not the role occupants'. The role expectations
in that case become the hypotheses that the researcher will test by his
research. Despite the problematic nature of role theory, its adoption gives
one a powerful tool for understanding and even generating expectations
about an individual actor's behavior (see Walker, 1987).
What sorts of roles can nations have? How are they to be defined? In
probably the most well known work relating roles to foreign policy (Holsti,
196 Valerie M. Hudson
1970), the author content-analyzed the speeches of leaders to identify

seventeen types of national role conceptions. Subsequent works in this area
(Wish, 1980; Walker, 1979) have relied on Holsti's basic typology. The national
role conception scheme of Holsti is based on two interesting assumptions:
first, that national role conception should be defined as the leader's role
conception for the nation and, second, that nations have a well-defined,
cross-time repertoire of roles. The answers could have been otherwise. For
instance, one could say that roles of nations can be analytical roles (in
Goffman's [1961] terminology)-that is, roles constructed by the analyst and
superimposed on the behavior of those examined. Second, one could answer
that there are two levels of national roles, the first representing the broader
ideological and instrumental beliefs of a national regime (corresponding to
Holsti's framework, as well as the concept of "regime orientation" found in
Hermann, 1986), the second being situationally defined roles that shift as
the situation does. The outcomes of roles in these situations make up or
shape the broader roles. As the reader may have guessed, I intend to pursue
these alternate answers.
To understand how role analysis is used in determining situational pre-
disposition, it is first necessary to comprehend a key assumption of the
model: that nations (governments and ruling parties) act in foreign affairs
only in response to a problem or opportunity they recognize. A problem
(or opportunity) can be seen as any perceived discrepancy between present
or anticipated states of affairs and what is envisioned as desirable. The
actor's definition of the problem sorts out who and what is important in
the situation. Vary the problem (or, more precisely, the definition of the
problem) and-naturally enough-from the perspective of the defining actor,
the relevant other entities are likely to vary, as are the interests, resources,
and needs that the actor finds applicabl~~
Every contemporary government is surrounded by numerous other actors
beyond its political jurisdiction-other national governments, intergovern-
mental organizations, and so on. The total number of collective actors runs
in the thousands. It is the nature of the particular problem that narrows
and defines the set of relevant other actors from the mass of all existing
foreign entities. For a government considering whether or not to support
the admission of a regime to the United Nations, the set of relevant others
will be quite different from those entering into the decision on how to
respond to a call for help from a nation suffering from the aftermath of a
natural disaster. Steinbrunner (1974) has suggested that in the cybernetic
approach to decision making, complexity is reduced to manageable levels
by focusing on only a few elements at one time. The concept of problem
provides a means for establishing a manageable set of external relevant
actors for governments.
Starting with the acting government's definition of the problem, it is

possible to establish three essential features of the external environment
that are powerful in explaining how an actor is likely to influence others
to cope with the problem. The first feature involves the identification of
basic roles in a problem and the entities that occupy those roles. These
roles, as mentioned above, will be defined situationally and by reference to
the actor's definition of the problem embedded in the situation. Knowledge
of the actor's role in the problem relative to the other roles enables one to
categorize the type of situation (the type of dramatic scene) being experienced.
The type of situation frames the task the actor will pose for itself. Using
information about the actual entities occupying all the basic roles, we can
examine the third feature, that is, the nature of the selective dimensions
of the relationships among the entities. Together these three elements make
possible the calculation of the situational predisposition for the actor-that
is, the probable configuration of foreign policy behavior properties that any
government would likely exhibit based on the dramatic requirements of its
role. This predisposition toward behavior can be, and very often is, altered
by the domestic features of the society and its political system, but those
parts of the larger model are beyond the concern of this chapter.
To illustrate the face validity of these features, let us take the case of
a theater fire (the reader is invited to draw parallels to an international
incident, as the following is on the individual, not the national level):
A fire breaks out in an empty theater. A person (the actor whose behavior
we wish to explain) discovers the fire and enters the theater. He sees another
person lying unconscious on the floor. What, if anything, will the actor do?
First, we can narrow down the range of behavior that will be appropriate
on the basis of the actor's role vis a vis the unconscious person. Some
types of behavior will be clearly inappropriate-like trying to buy a pack
of bubble gum from the unconscious person. The other person, from the
actor's definition of the problem, is both the source of the problem (he is
unable to move away from danger) and the subject of the problem (because
he will likely die if he does not move). Thus, the behavioral question facing
the actor is: Should I give assistance to the unconscious person? The actor's
answer will be based on his assessment of the situation and his relationship
to the unconscious person. Our actor could be a small child, and the other
a large, heavy-set man. Does the actor have any motivation to help? For
instance, is the other person a friend, an enemy, a stranger? Are there any
other sources of assistance-did the actor hear a fire engine pull up outside?
We can reset the entire scenario by resetting the actor's definition of the
problem and the role he is playing in the situation. Let us now suppose
our actor is no innocent bystander but an arsonist who has set the fire.
He enters the theater to view his handiwork and confronts a very conscious
policeman. The type of situation has completely changed for our actor.
Now, from the actor's perspective, the source of the problem is the policeman,
and the actor himself is the subject because he could be apprehended.
Knowing the actor's role vis a vis the policeman, we can see the situation
is one of confrontation in which the actor's decision on behavior is driven
by the task: How can I reduce the adverse effects the policeman is causing
for me? How can I get out of here without being caught? If the actor is
an accomplished enough liar, he could try and bluff his way out of the
situation by pretending to be an innocent bystander. Depending on his
assessment of his chances with regard to the policeman, he could try to
overpower him or run. Other aspects of the relationship may be important-
what if the policeman is the actor's father?
These variations have been designed to suggest the plausibility of roles,
types of situations, and relationships (taken in the context of a specific
problem perceived by an actor) in shaping influence attempts.
Knowing the actor's situation defines his task and his relevant audience.
According to role theory, knowledge of an individual's task and the relevant
others establishes the role the individual can be expected to assume.
Furthermore, and where role theory presents itself most powerfully, "role
may be defined as the typical response of individuals in a particular position"
(Goffman 1961: 93). Analytical role theory is used to gain insight into and
expectations about typical behavior of an entity, without having to know
more about that entity's individual attributes than its relative position in
some defined role system. To apply this form of explanation, it is necessary
to specify the particular kinds of roles, situations, and relational dimensions
that affect behavior. That discussion follows.
Roles in the Power Drama

From the perspective of the acting regime or leader (the actor), every
recognized problem will produce a role set that includes the actor role, the
source role, and the subject role. 2 The key role in any problem is that of
the actor. The actor's perspective defines the identity of the other two role
occupants, thus indirectly determining the type of situation. In this model,
the acting nation's perspective is equated with that of the individual policy-
makers who have the authority to allocate resources to foreign policy. The
analyst selects an entity as the actor in the situation because it is the entity
whose behavior is to be explained.
In addition to the actor, every problem has an inferred source or a
"cause." The "source" is that entity(ies) which is perceived by the actor
to be responsible for the immediate problem (regardless of whether such
perceptions are accurate). Any human entities-governmental or nongov-
ernmental-or nature or the will of God may be regarded as a source by
the actor's policy-makers. Every problem also has a perceived "subject."
The subject is the entity(ies) perceived by the actor to be most directly

affected by the problem in the immediate situation.
Each role may be occupied by one or more separate entities. The same
entity may also appear in several roles. The entity in the source role may
also be in the subject role. For example, the actor may also be the subject,
the source, or both. 3
This identification of actor, source, and subject in a discrete circumstance
will be the basis for the two major sources of information to be gleaned
from the situational context. The specific configuration of the occupants in
the roles will define the type of situation: The relationships between the
actor and each of the other primary entities in the situation will provide
the relational network. The first frames the task of choice facing the actor;
the second summarizes the particular slice of the broader social system
within which the role behavior will be enacted. The type of situation will
outline the general goal or strategy, and the relational network will add more
specific motivations, capabilities, and constraints. With these sources of
information in one's possession, it is possible to attempt the challenge of
estimating the actor's foreign policy behavior in a discrete case.
The model utilizes the following fivefold typology of situations (assume
the following international entities: the actor A, nation X, and nation V):
I. Confrontation
Diagram: Actor Source Subject
A X A
or
A A X
Explanation: In this type of situation, the actor also occupies one of the
other primary roles, usually the subject role. The problem caused by the
source directly affects the actor's enjoyment of its basic values.
Choice Facing the Actor: How can we reduce the adverse effects that
the source in the problem creates for us?
11. Intervention
A X V
Explanation: In this situation, the actor plays neither the source nor
subject roles, which in turn are occupied by different entities (i.e., the source
and the subject entities are not the same). A problem for V has been caused
by X, and the actor (A) is not directly affected.
Choice Facing the Actor: Should we intervene in this problem on one
side or the other and, if so, in what manner?
200 Va/erie M. Hudson
Ill. Assistance Request

Diagram: Actor Source Subject Secondary Role
A A A y
Explanation: Here the actor plays all the major roles, with another entity
playing a secondary role. (If there were no other entity, it wouldn't be a
foreign policy problem!) As mentioned earlier, this is the international
equivalent of "shooting oneself in the foot" -i.e., the actor is experiencing
some domestic problem and seeks assistance.
Choice Facing the Actor: Who can give us assistance in reducing the
adverse effects of the problem we are experiencing?
IV. Assistance Consideration
A X X
Explanation: This situation is the complement of the previous one, with

another entity occupying both source and subject roles. Entity X is facing
a problem and seeks assistance from the actor.
Choice Facing the Actor: Should we provide assistance to those who are
experiencing the adverse effects of a problem?
V. Collaboration
A A+X+Y A+X+Y
Explanation: Here the actor and other international entities are all jointly
facing a problem. Together they occupy both source and subject roles.
Their joint activity, best described as negotiation or collaboration, will be
an attempt to address their mutual problem.
Choice Facing the Actor: Can we reach a substantive agreement with
those with whom we share this problem?
Actor Motivation and Capabilities

The concept of "relational network" can be viewed as capturing the
motivation of the actor in the dramatic scene and also the capabilities the
actor possesses and can use in the scene. Briefly put, the model examines
three dimensions along which the relationships between the actor and each
of the other role occupants can be assessed: prior affect, relative capabilities,
and salience. Prior affect establishes the affective history between the actor
and each of the other entities. Affect is expressed feeling toward another
entity ranging from extreme hostility to unequivocal friendship. The ac-
cumulated manifestations of affect in previous interactions that the actor
has expressed toward a particular entity compose the prior affect between
the two, from the actor's standpoint. It also provides motivation for the
actor to express positive affect toward other entities that treat the actor as
he would be treated, and negative affect toward those who do not.
Relative capabilities involve a calculation of another entity's capability in
various areas vis a vis that of the actor. The areas of comparison can be
classified according to the basic value to which they are relevant: security I
physical safety, economic wealth, respect/ status, and well-being/ enlighten-
ment.4 Underlying this comparison of capabilities is the actor's desire to
consider whether it has capabilities sufficient for whatever undertaking it
plans, and whether the other entity has a level of capabilities that would
neutralize the effect or lead to retaliation.
Salience for an actor is the degree to which the explicit or tacit support
or concurrence of a specific external entity is necessary for the regime, or
its society, to realize basic values. In other words, the well-being of one
nation is contingent upon the resources or agreement of another. A hostile
entity can also be viewed as salient, as the immediate fortunes of the actor
may be highly dependent on an enemy's actions or inactions. Salience often
acts as a brake or constraint on the motivation provided by prior affect,
especially when negative action is contemplated and may be a reinforcement
for positive steps.
With the concepts of "role," "type of situation," and "relational network"
defined and placed within a theoretical context, the explanatory framework
for this situational predisposition model is complete. A word must be said,
though, about the dependent variable, foreign policy behavior, and its
conceptualization, before proceeding to the second task of this paper
(methodological choices in the construction of the model). Foreign policy
is fundamentally an act of communication, as it represents an attempt to
influence others. lt is useful to think of foreign policy behavior as representing
different properties of any communicative act-who does or conveys what,
to whom, how. (See Callahan et al., 1982.)
For the moment, regard the actor (i.e., "who") as a given-that is, what
national government or ruling party will be the actor is specified in advance.
The behavioral properties the model should explain are (1) the recipients
(to whom?)-the nature of the entity(ies) the actor will address; (2) the
affect (does or conveys what?)-the manifest feelings of approval or dis-
approval the actor expresses; (3) the commitment (does or conveys what?)-
the indication of resolve or degree of binding itself that the actor conveys;
and (4) the instruments (how?)-the skills and resources of statecraft that
the actor uses. Each of these four behavior properties is elaborated in
Callahan et al. (1982).
These measurable properties are common to all foreign policy behavior.
If we are able to understand why they are likely to assume certain values
under certain conditions, we have gained much that is of practical and
theoretical worth in understanding foreign policy. 5 It is these foreign policy

behavior properties that the situational predisposition model seeks to explain.
Methodology and Model Construction

The nature of the conceptual framework poses some stiff challenges,
methodologically speaking. To construct a testable model, the following
points must be recognized:
1. The explanatory variables are qualitative in nature.

2. The dependent variables are also basically qualitative-only affect and
commitment can be placed at the ordinal level of measurement precision.
3. The variables do not interact or interrelate in a simple or consistent
manner. (For example, in certain types of situations, certain relational
dimensions become more important, whereas some may become
irrelevant.)
4. The aim of the model is to estimate or predict discrete instances of
foreign policy behavior, hence time cannot be viewed as continuous.
5. The number of possible combinations of all the values of each of the
explanatory variables is very large: The methodology chosen must be
able to utilize to the fullest extent this rich and complex knowledge
base.
As there is no statistical or mathematical method able to accept all these

conditions, it is a natural impulse to suggest that the only means by which
the framework can be implemented is through the case study approach.
This is indeed one option but does not readily afford us the capability of
cross-national aggregate analysis. The case study approach would not allow
us to explore easily all of the possible combinations of variables, without
which it would be difficult to gauge the efficacy of the situational predisposition
model as a whole. Furthermore, because this is designed to be a general-
use model, it would be helpful to apply and test it in a generalized, not an
idiosyncratic, manner.
The alternative approach chosen is that of "artificial intelligence-type"
modeling. The reason the term "At-type modeling" is used is that it is
important to keep in mind that what is spoken of here is a broad approach
within which numerous, diverse, more narrowly defined methodologies may
be found. These methodologies all share some common features: They take
a "human intelligence" approach to problem solving and explanation. 6
Heuristics7 are used to organize, interrelate, and subsequently search large,
complex knowledge bases. Indeed, this type of knowledge base is another
hallmark of the AI approach to modeling. Last, these types of models are
almost always instantiated as computer programs.
Al-type modeling is, as a consequence, very flexible-much more so

than most formal methods used in international relations to date. The models
can be parsimonious or horrendously complex-no preference underlies the
approach. The data used can be at any level of measurement precision,
and there are no restrictions on the number of variables employed. The
rules by which the variables are interrelated are limited only by the bounds
of the individual researcher's creativity, as long as they are not logically
inconsistent. The only thing an Al-type model requires is that the researcher
be as explicit and complete as possible.
The particular method used to construct (and ultimately test) the situational
predisposition model is the rule-based production system. Most rule-based
production systems operate as follows: A set of initial conditions (IC11 ) is
programmed into the system, and then according to a set of rules created
by the researcher (or procedures deducible from those rules, either being
denoted by the symbol of the arrow, - ), the system assesses the effects
of inputs (explanatory variables, X) on the initial condition (IC11 X X
-ICX11 ), postulates the value of the outputs (dependent variable, Y as
follows: ICX11 -V), and updates the initial condition (Y - I C12). As
Winston (1977: 144) points out, the rules in a production system can be
seen as linking situations with actions or, in other systems, linking premises
with conclusions.
The rule-based production system used to implement the situational
predisposition framework is constructed as follows:
IC11 : The system program is given the initial conditions, these corresponding
to information on the relationships between each actor nation and every
other international entity of interest. The relational network data is coded
and stored along the three aforementioned dimensions of prior affect, relative
capabilities, and salience. The information will serve as a backdrop against
which the system will evaluate the discrete situations provided by X.
X: The inputs to the system are provided from external sources-i.e., the
user of the system. Information is supplied on specific situations facing a
particular actor from that actor's perspective. This information identifies the
actor, the other relevant entities involved in the situation, and also which
of several roles they occupy: actor, source, subject, facilitator, aggravator,
etc. The basic values at issue in the situation (security, economic wealth,
respect/status, etc.) are also supplied.
--.: There are two distinct sets of rules at the heart of the situational
predisposition model. The first set of rules, described earlier, specifies what
type of situation exists (e.g., confrontation, intervention, collaboration, etc.),
given the configuration of roles and occupants provided by X. These types
of situation will frame the task facing the actor nation and serve to narrow
down the range of probable and/ or appropriate behavior from the set of
all possible national behaviors. The second set of rules specifies how knowledge
of the situation type and knowledge of the relationships between the actor
and the other entities (as given in IC) are to be combined. Please refer to
Table 8.1 for clarification.
In the model, the second set of rules takes the form of extended decision
trees utilizing three subtypes of rules, that (1) postulate the outputs given
the variables in isolation, (2) then postulate the outputs as the result of the
interaction of the explanatory variables (3) according to rules that govern
the form of the interaction. These trees are rather complex, with there being
221 of the behavior rules alone, which do not include "isolation" rules or
interaction rules. (Refer to Table 8.1 for further explication.)
All rules are provided the system by an "expert"-in this case, the
research team. They represent our own hypotheses concerning foreign policy
behavior (see Hudson, 1983; Hudson, Hermann, and Singer, 1985). To
construct the rules, the effects on behavior of each value of the explanatory
variables (prior affect, relative capabilities, and salience) in each type of
situation were postulated. These are called the "isolation" rules because no
interaction of the explanatory variables is considered; e.g., prior affect values
are examined in isolation from relative capability and salience values.
Next, interaction rules govern which variables would take precedence
over others in various situations given the values of those variables. An
example is provided in Table 8.1. These interaction rules can be characterized
variously as (a) additive interaction rules, where the expected values from
the isolation rules are assigned equal weight and then averaged in an
arithmetic fashion, (b) override interaction rules, in which one expected
value is postulated to dampen the effects of the others, or (c) relevancy
interaction rules, where some variables values are considered irrelevant given
the type of situation and the values of the other explanatory variables.
Finally, the behavior rules, where each combination of explanatory variables
values in each situation was examined and a determination of resultant
behavior made, were created by use of the isolation rules as informed by
the interaction rules. Each branch of this final, interactive, immense decision
tree is thus an empirically testable hypothesis that indicates how different
values of the explanatory variables interact to produce an expected set of
foreign policy behavior properties.
V: The output of the second set of rules will be an estimate of the actor
nation's behavioral response to the discrete situation, given in terms of an
expressed level of affect, level of commitment, choice of instrumentality,
and identity of the recipient of the behavior.
Y, the output, can then be used to update IC, the relationships between
the actor and the other entities involved. The production system will store
the affect expressed to the recipient as the latest in the time series of prior
affect points between the actor and the nation/entity that is the recipient.
Table 8.1: Types of Rules in the Situational Predisposition Model
I. Roles/Situation Rules
These rules specify what type of situation exists given the configuration of
roles in a given case.
EXAMPLE:
If the ACTOR plays neither the SOURCE nor SUBJECT roles, and the occupants
of these two roles are NOT the same, the SITUATION TYPE is INTERVENTION.
11. Explanatory VariablefBehavior Attribute Rules

These rules specify what the value of the behavior attributes will be for the
actor in a specific case, given knowledge of all the explanatory variables-
situation type, prior affect, relative capabilities, salience, and presence/absence
of facilitators.
There are three levels of explanatory variablefbehavior attribute rules, with
the third and final level being the comprehensive set used in the empirical testing.
They are:
Level A: "Isolation Rules"
These rules specify the effects of each of the possible values of the relational
dimensions (prior affect, relative capabilities, and salience) in isolation from the
others, in every conceivable type of situation.
EXAMPLE:
If the SITUATION TYPE is CONFRONTATION, and the SOURCE possesses a
significant CAPABILITY ADVANTAGE over the ACTOR, the actor will most likely
respond using DIPLOMATIC INSTRUMENTS and will NOT use HIGH COMMIT-
MENT.
Level B: Interaction Rules
These abstract rules specify how various combinations of the values of the
explanatory variables are likely to interact.
EXAMPLE:
In an ASSISTANCE CONSIDERATION SITUATION, PRIOR AFFECT takes over-
whelming [RECEDENCE over the values of the other relational dimensions. In
fact, in this type of situation, RELATIVE CAPABILITIES need not be calculated,
unless it be those of any facilitators when PRIOR AFFECT has been NEGATIVE
or MIXED.
EXAMPLE:
In general, if another role occupant is SALIENT for the ACTOR, this will tend
to DAMPEN the INTENSITY of NEGATIVE AFFECT overall, and the LEVEL of
COMMITMENT in COFNRONTATION situations as predicted by the "isolation
rules."
Level C: Behavior Rules
These rules specify the values of the behavior attributes for every possible
combination of the explanatory variables. These rules are based upon the
"isolation" and "interaction" rules.
EXAMPLE:
In a CONFRONTATION SITUATION, if PRIOR AFFECT between the ACTOR and
the SOURCE has been NEGATIVE, and this other entity is SALIENT for the
actor, and the actor's RELATIVE CAPABILITIES are greater than this other entity's,
and there are NO FACILITATORS/AGGRAVATORS present, the likely behavior
attribute values for the actor will be HIGH NEGATIVE AFFECT, MODERATE
COMMITMENT, DIPLOMATIC AND RELEVANT DOMAIN INSTRUMENTS, with
the SOURCE as RECIPIENT.
Figure 8.1: Excerpt from the Situational Predisposition Model Decision Logics
(2034) (2035)
(1005)
The updated IC will then be used to evaluate any new X that is supplied
to the system.
This system can be used for purposes of explanation of single cases or
for more aggregated tasks of postdictive testing of the model or even for
policy prediction itself. It might be useful at this point to examine a subsection
of the overall decision tree to illustrate how the system might work. Figure
8.1 is an abridged version of the decision tree for intervention type situations.
lt represents in modeling terms ICX-- V for this particular situation's
values of the ICX. lt is impossible to reproduce on one (or even ten) pages
the multiplicity of paths through this tree. 8 Six particular branches are
shown in full, while the other 100 or so are only implied by the diagram.
However, if one examines the decision node points of the six branches that
are shown, it should be clear what the remaining branches must look like.
Figure 8.1 is to be understood as follows: In an intervention type situation,
the actor is neither the subject nor the source, and the entities occupying
the source role and the subject role are not the same. The actor must
decide whether it will choose to side with either the subject or the source
and, if so, in what fashion. The first important piece of relational information
to the actor is what the prior affect has been between itself and these two
other role occupants. The ten possibilities represent the first level of decision
points on the tree (e.g., Positive/Positive-prior affect to both the source
and subject has been positive; Positive/Negative-prior affect to one has
~
~
(lD48)
been positive and to the other negative; Negative/ Mixed; Neutral-None/

Mixed, etc.) This first level activates the isolation rules for prior affect,
which suggest some broad maxims about what will be likely behavior and
what will not. In one case, where prior affect to both the source and the
subject have been negative (branch number 2041), an interaction rule kicks
in to determine that in such a case, no further relational knowledge is
needed, and a behavioral prediction is made on the basis of prior affect
alone.
In the other five branches shown in full, more relational information is
needed before an estimate can be made. The second level of decision points
concerns salience. Specifically, in an intervention situation, it would be
desirable to know whether both the subject and the source are salient to
the actor, if neither is, or if one is (and which entity that happens to be).
Again, this information brings to bear the isolation rules for salience and
the behavioral expectations that flow from those rules.
In each of these five cases is it deemed necessary to determine relative
capabilities, so the third level of decision points concerns that variable. The
information to be collected here is whether the actor possesses a capability
advantage in relation to either the source or the subject in the area of
capability deemed relevant by reference to the basic value threatened by
the problem in the situation. The actor could have an advantage over both,
over neither, or over either the subject or the source. In some cases, it is
desirable to unpack this knowledge even further-for branches numbered

2034 and 2035, where the actor is deemed to have no capability advantage,
it was important to know whether the entity toward which the actor has
expressed negative prior affect has a facilitator whose strength greatly
augments what that negative entity might otherwise have at its disposal.
With the expectations of the isolation rules for relative capabilities added
to those of prior affect and salience, enlightened application of the interaction
rules permits the combination of all the variables, resulting in a unique
behavioral estimate for each branch of the decision tree. Each branch is a
distinct, testable hypothesis and is the embodiment of the particular inter-
action rule used to create it. In other words, the situational predisposition
model contains 221 specific hypotheses linking the values of the explanatory
variables to values on the dependent variables.
Let's take the case of the 1956 invasion of Hungary by the Soviet Onion.
What would the United States do in reaction according to the model?
Hypothesis (or branch) number 2048 in Figure 8.1 is arguably applicable.
(See identifying numbers at the end of each branch in the figure.) The
United States perceives the USSR has created a problem for Hungary and
must decide whether to become involved. Hence, this is an intervention
type situation for the United States. Prior affect between the United States
and the USSR is negative; prior affect between the United States and Hungary
is more neutral. The USSR is salient for the United States; Hungary is not.
The United States possesses a capability advantage over Hungary in the
area of military capabilities, but not over the USSR. The behavior rule for
this combination of values of the explanatory variables predicts the United
States will sharply rebuke the USSR, but without high commitment of its
own resources (high negative affect, minimal commitment, diplomatic in-
strument, address negative entity).
One advantage of this model is that we can easily estimate the differences
in the United States' behavioral predisposition as the particulars of the
situation are varied (counterfactual analysis). Assume now the United States
and the USSR over future decades have become longtime close allies, but
everything else is the same as in the previous example. How does the O.S.
response differ? According to the branch in Figure 8.1 labeled 2088, now
the United States will actually express low-level support of the OSSR's
invasion (low positive affect, low commitment, diplomatic instrument, address
positive entity). What if not only the United States and the USSR are allies
but the United States and Hungary are allies, and both are salient for the
United States? Here, especially if the United States has a capability advantage,
it will attempt to mediate between the two sides (hypothesis/branch number
2005: low positive affect, moderate commitment, diplomatic instrument,
address both source and subject entities).
In sum, the rule-based production system approach as a methodological

choice for implementing this model offers some clear advantages. It has
not only provided a flexible yet systematic way to link qualitative explanatory
variables to qualitative dependent variables but, as we shall see, is also an
excellent vehicle for testing the accuracy and sensitivity of this forecasting
model. An initial test follows.
Empirically Testing the Model
Constructing a Test
Before we can evaluate the situational predisposition approach, and answer
the "larger" questions posed in the introductory section (e.g., whether it is
possible to incorporate notions of a dramatic script into foreign policy
analysis), the model must be subjected to an empirical test.
Because the rule-based production system is already in testable form
(i.e., all hypotheses have been explicitly and comprehensively outlined), the
first step is the operationalization of the variables. Specifically, operational
measures were constructed for the identification of role occupants in a series
of discrete situations, the relational network variables (prior affect, relative
capabilities, and salience), and the foreign policy behavior attributes (affect,
commitment, instrumentality, and recipient). An event data set that provided
information on discrete situations was used, resulting in a sample of 6605
cases for twenty-eight actor nations during the decade 1959-1968. 9 The
data set chosen contains a number of foreign policy behavior properties for
each event, including the ones used as dependent variables in the situational
predisposition model. (See Callahan et al., 1982.)
Subsequently, the same events were recorded to specify the perceived
problem and the role occupant designation from the actor's perspective. A
description of the circumstances that existed prior to each event established
the basis for judgment as to which real international entities should be
assigned to the roles in each problem. 10 Two judges coded all problem
situations and reached agreement on the assignment of roles (the inter-
coder reliability on the role coding ranged from .80-.85). As explicated
earlier, that role coding automatically determines the type of situation because,
as diagrammed in the section on the conceptual framework, the configuration
of roles defines the situation type.
Data sources other than the event data set discussed above provided the
measures for the relational network variables. The operational procedures
for these variables are described in full in the Situational Predisposition
Operationalization Manual. 11 The sources utilized included, among others,
the Conflict and Peace Data Bank (COPDAB), International Monetary Fund
and IBRD Direction of Trade statistics, Keesing's Contemporary Archives,
the USACDA's World Military Expenditures and Arms Transfers, and the
United Nations Statistical Yearbooks.
The test itself was conducted as follows: I "fed" the system the operational
data inputs and checked its outputs against the "reality" of what the nations
did. For each of the 6605 cases, the determination as to which hypothesis
(branch) was applicable was made, using the data on situation type and
the relational network. The behavior predicted by the model in each case
was compared with that recorded as having actually occurred in that discrete
instance by the event data set. Although the estimated behavior properties
are referred to as those "predicted" by the model, it should be acknowledged
that this test is technically one of postdiction rather than prediction. The
same procedure, however, could be used in a predictive test if future event
data were collected after the model estimates were made.
Results
The focus of a test is the model's outcome validity. 12 Does it produce
outputs that agree with the outputs of reality? 13 Some thorny issues face
those wishing to evaluate their models on the basis of empirical accuracy,
though. 14 The most pressing concern is how to regard the results one obtains
by such a test. If the returns do not yield 100 percent accuracy, does one
dismiss the model? Such a standard is unrealistic. There may be slippage
because one could not include all variables, because one's data were not
free from error, etc. But where to draw the line is not clear. When does
one conclude there is something definitely amiss with one's system? Below
70 percent? Fifty percent? There is no easy answer.
In practice, one may have to supplement measures of accuracy with
other measures of conceptual validity or predictive utility. In the research
presented here, the model constructed purposefully excludes several important
classes of variables (such as domestic and systemic sources of foreign
policy behavior). As a result, we would not expect accuracies as high as
if these other elements were included. At this point, it is useful to determine
whether the variables in the model have predictive capability and make
meaningful distinctions between conceptually different types of foreign policy
situations. lt is important to recognize that the model could produce significant
situational differentiation, even if the rules linking those distinctions to
behavior were off somewhat (resulting in inaccurate predictions). This assertion
could not be tested, obviously, with simple measures of postdictive accuracy.
Therefore, two means of evaluating the results of the empirical test are
used. The first approach assesses the model's predictive accuracy by
ascertaining what percentage of the time the predictions were correct (simple
predictive accuracy). In this approach, the percentage of correct predictions
for each of the individual behavior attributes (affect, commitment, instru-
Table 8.2: Predictive Success for Separate Behavior Properties
Property \ Correct Phi*

Affect 49\ .43
Commitment 24\ .28
Instrument 62\ .39**

Recipient 63\ N/A***
* = Chi-square results were all significant to the .001 level.
** = Phi for this variable is artificially low because of the
extreme skewedness of the distribution.
*** = For coding reasons no meaningful two-way analysis of this
variable could be constructed.
mentality, and recipient) found in the nation's actual response to the situation
is noted, as well as for the combinations of those behavior attributes predicted
by each branch/hypothesis in the decision tree.
The second approach (organizational performance) assesses the model's
utility for forecasting by focusing at the level of the individual hypotheses
rather than at the aggregate level. Here, not only can we ask if a specific
hypothesis generates accurate predictions but we can also determine support
for the hypothesis in a more basic sense. We can analyze whether each
specific hypothesis (branch) of the decision tree tends to result in unique
patterns of actual (rather than predicted) behavior. This second type of
support indicates if the model's explanatory variables differentiate patterns
of behavior in a meaningful fashion, whether or not the model predicts it.
Simple Predictive Accuracy. Table 8.2 presents success rates for predicting
each of the four behavior properties separately. The model does well in
predicting recipient (63 percent correct) and instrument (62 percent). Lower
prediction rates occur with affect (49 percent) and with commitment (24
percent). Some of the explanation for the lower success rate with affect
may be its partial dependence on first having to correctly establish the
recipient in the situation. Because affect is conveyed to a specific type of
recipient, any error in predicting recipient means the model is more likely
to be wrong about the affect. When the recipient is correctly predicted,
affect success rates improve (58 percent correct).
In the conceptualization used by the model, any foreign policy behavior
combines the four properties that together establish who was addressed
(recipient), how action was taken (instrument), and what was done or
Table 8.3: Correct Predictions of Combinations of Behavior Properties
(N=6605)
Number of
Pro12erties Correct ' of Sample Cumulative
'
4 6.0 6.0
3 23.9 29.9
2 37.5 67.4
1 26.5 93.9
0 6.0 100.0
conveyed (affect and commitment). Accordingly, it is useful to determine

how well the model performed in predicting the combination of the four
properties. Table 8.3 shows the success rate for the 6605 cases. The
predictions for each case could be completely wrong (none of the four
properties correctly predicted), completely right (four of the four properties
correctly predicted), or somewhere in between. As Table 8.3 shows, the
situational predisposition model's predictions of the combined effects were
accurate 30 percent of the time for three or four out of the four behavior
properties. Less stringently, the model correctly predicted two or more (two,
three, or four) in 67 percent of the cases. The model fails to predict at
least two of the four properties (zero or one) about 32.5 percent of the
time. The clear implication of these numbers is that the situational predis-
position is most formative in about one-third of the cases examined, somewhat
formative in another third, and not very formative in still another third.
This test allows us to speculate about the types of conditions under which
the situational predisposition has considerable impact and those where the
other kinds of variables (e.g., domestic or systemic factors) may be more
important in shaping behavior.
Organizational Performance. This analysis is limited to 47 of the 221
hypotheses or branches. Each of these 47 hypotheses had 25 or more cases
from our sample-the number judged to be of sufficient size to enable the
researcher to recognize patterns of behavior and to draw plausible initial
generalizations. Together these 47 captured 80 percent of the 6605 events
being studied.
Before beginning this discussion, we should recall that as a result of this
empirical test, not only does each branch have a prediction of behavior but
each branch now also has a distribution of actual behavior properties, i.e.,
the model has sorted the actual behavior of the twenty-eight nations as
recorded by the events data set according to the particular applicable branch.
It is to these distributions of actual behavior properties that I will now make

reference.
When actual events are distributed among the selected hypotheses or
branches, it is useful to know two things. First, do behaviors with similar
properties tend to occur more frequently than others in a given branch?
Second, did the model predict what that cluster of behavior properties would
be? These distributional clusters, if they exist, lend credence to the assumption
that the model's explanatory variables are important in differentiating kinds
of foreign policy behavior. If one kind of behavior dominates a given decision
tree branch, whereas other behaviors prevail in other branches, it suggests
that the model has incorporated a small set of powerful variables that are
in fact making discriminations that are conceptually and postdictively (and
by implication, predictively) useful. Furthermore, if the prediction for such
a branch matches these behavioral patterns, then there is substantiation
not only for the significance of the variables but also for the model's
interpretation of how they operate. Of course, if the model fails to predict
the prevailing cluster of behavior, then subsequently the explanations regarding
the reasons and conditions governing the observed pattern must be recon-
structed. Even so, by the very fact that a pattern exists, the model will
have taken a significant step toward isolating variables and interaction rules
that are associated with particular kinds of foreign policy.
To determine if clusters of foreign policy behavior with similar properties
dominated a given branch, I noted the difference in percent between the
most frequent category (or "peak") of each behavior property and the next
most frequent category. Three successively larger thresholds of difference
between the peak and second most common category were examined. They
were: a difference between the peak category and the next highest of greater
than 10 percent; of greater than 20 percent; and of greater than 30 percent.
The presence of a peak greater than the 30 percent threshold indicates a
very substantial clustering of behavior around a particular category. If the
model's prediction of the behavior property for a branch matched the
category of the prevalent peak, then the model achieved success. This
phenomenon is termed a "match."
Table 8.4 identifies both the presence/absence of distributional peaks
and predicted matches. If every one of the distributions of the four behavior
properties in the forty-seven branches had a peak category, then there
would have been 47 X 4 = 188 peaks to be observed. It is encouraging
to note that peaks of behavior were discovered in 162 distributions (86
percent) at the 10 percent or above threshold. They occurred 62 percent
of the time at the 20 percent or higher threshold and 48 percent of the
time at the most stringent threshold (30 percent or higher). It would appear
that the model's variables sort out different kinds of behavior rather well.
Table 8.4: Number of Distributional Peaks of Frequent Behavior and Percent of Those
Behaviors Predicted
Differences of 10\ or more
• of distributional peaks at this threshold 162 (86%) 2
' of matches (successful predictions)

94 (58\)
11 of peaks at this threshold 117 (62\) 2
• of matches
77 (66\)
11 of peaks at this threshold 91 (48\)2
11 of matches 65 (71\)
1 The 47 branches (hypotheses) of the decision tree reported in

this table are those for which 25 or more actual cases of foreign
policy occurred in the data set of 6605 cases.
2 Percentage based on 47 x 4 = 188 possible opportunities for.
Were the peaks of frequently occurring categories of behavior those which

the model predicted? Of the 162 distributions that had a single most
frequently occurring category, the model predicted eighty-nine, that is, 58
percent. Table 8.4 also shows the number of matches at the higher peak
thresholds. Note that the more prevalent the category is in a distribution
(i.e., the higher the threshold), the greater percentage of matched predictions.
Thus, at the highest threshold, 71 percent of the most frequent behavior
was predicted. This is as it should be. The more likely a branch is to result
in one particular category of a behavior property, the more frequently the
model should be able to discern it. The success in predicting the most
commonly occurring categories of behavior is most encouraging. However,
by all rights the percentage of matches should be close to 100 percent. If
the decision trees are in fact revealing peaks 62 percent (or whatever
percent) of the time, then the model ought to be able to project every
single one of those peaks.
The upshot of this second type of analysis is to show how simple
measures of predictive accuracy may need to be "unpacked" before the
strengths and weaknesses of a particular system can be discovered. This,
in turn, leads us to ask what is the most appropriate way to modify and
refine a rule-based production system once weaknesses have been discovered.
The previous discussion's identification of "missed matches" (i.e., where
peaks occurred but the system failed to predict them) provides a handy
starting place. A few guidelines should be established first.
Clearly, one impulse to be suppressed is that of rampant ex post ante-
type tinkering. If the model is bludgeoned into being able to postdict perfectly
to the sample its accuracy was initially tested by, it will probably have been
robbed of any real power it had, either as a predictor or as a theory.
Tinkering must be done in an enlightened manner, through the use of such
techniques as the identification of "missed matches." Still, not everything
suspected of being an error or a weak point should be immediately altered.
That aspect of the model may simply appear weak as a consequence of
the small slice of "reality" against which it has been tested. A good rule
of thumb for tinkering may be, if you can't theoretically justify the change,
don't make it until you can. Perhaps more testing will afford a clearer
perspective.
Conclusion
Governments, this model argues, sort out problems according to their
perception of the dramatic requirements of the situation at hand. The
situation frames the problem by establishing who is involved and how they
are related to the acting government. Knowledge of the script of the power
drama and the actor's role in the drama leads us to project the major
properties of the actor's foreign policy behavior. How well does the script
help us predict the actor's behavior? How well can we account for a
government's foreign policy without taking into account domestic factors
and consideration of the requirements of a particular political system?
The initial test suggests that the situational predisposition model can
account for an appreciable amount of foreign policy activity. Such central
characteristics of foreign policy behavior as the nature of the recipient and
the type of instrument that will be used can be correctly forecasted over
60 percent of the time, and, where recipient is accurately predicted, the
affect expressed to that entity is correct in 58 percent of the cases. The
24 percent success rate in forecasting the fourth property-commitment-
needs improvement. That seems possible given the emergence in the model's
initial results of distinguishable patterns of behavior, including patterns in
the use of commitment.
Of the most frequently applied hypotheses (branches) of the model in
the test, 86 percent were dominated by a distinctive pattern of behavior. 15
These configurations of dominant behavior were predicted 58 percent of
the time by the model's hypotheses. This rather high rate of support for
the hypotheses may be the most important finding of all. The hypotheses
are the embodiment of the three-tiered set of decision rules that postulate
how situation type and relational variables combine to form a decision
calculus in a particular case. That so many events for twenty-eight diverse

nations are accounted for by the combination of the model's explanatory
variables seems to indicate a promising source of explanation that cannot
be neglected. Furthermore, substantial clustering of certain kinds of behavior
that the situational model produced but were not accurately predicted by
the hypotheses of the model offers information on how its future performance
might be sharpened. How much the model's operation can be enhanced in
this fashion remains a question for the future, but there are probably
significant limits on how much better it can perform alone.
The reason for this expectation is simple: The model is intended to
generate the behavioral predisposition of any government in response to
the dramatic requirements of specific kinds of situations. In other words,
the model assumes that all regimes will respond in a similar way when
faced with a comparable dramatic situation, as defined by the explanatory
variables. The results of the test show, and common sense confirms, that
this is not always the case. As a result of particular societal conditions,
domestic political requirements, or the personal characteristics of political
leaders, different governments are often likely to respond to the same type
of situation differently. A more elaborate model of foreign policy that includes
key domestic factors is needed. The situational predisposition model dem-
onstrates that the extent to which nations will react similarly in similar
circumstances because of the script of the international power drama can
be more precisely calculated. The model also suggests that in any more
extensive undertaking, a component based on situational or dramatic logic
will be an essential and powerful contributor.
Notes
1. The model is fully outlined in Hudson (1983), and variants of it have appeared
in papers co-authored by myself, Charles F. Hermann, and Eric G. Singer (see, for
example, Hudson, Hermann, and Singer, 1985). Professor Hermann was instrumental
in developing the isolation rules of the model, and the genesis of the role categories
can be traced to Hermann and Hermann (1979).
2. There are secondary roles as well, e.g., facilitator, aggravator, potential facilitator,
etc., but to keep the discussion brief, these will not be examined. See Hudson,
1983, for an explication of these secondary roles.
3. The latter situation is analogous to an individual deciding to treat his own
injury after shooting himself in the foot. This international equivalent of "shooting
oneself in the foot" would not be a foreign policy occasion unless there were some
other role involving a foreign entity in addition to those mentioned. A secondary
role, such as facilitator (see note 2) must be present.
4. These basic values are drawn from a more comprehensive set proposed by
Lasswell (1971). The model uses, with some combination, those judged to be common
in international affairs.
5. Of the numerous possible ways of conceptualizing the elements of foreign

policy behavior, a strong case can be made for the merits of a scheme that (1)
can be observed in a reliable fashion, (2) can be found in all instances of the foreign
activity of any actor, and (3) can describe behavior in terms of widely recognized
features of communication, including target (recipient), disposition (affect), deter-
mination (commitment), and means (instrument). The utility of these four behavior
properties in interpreting foreign policy is elaborated in Callahan et al., 1982.
6. This does not mean that an Al-type model can only be used in cognitive
studies or the like, although it is true that this approach has so far been used by
those in foreign policy studies who have also adopted the decision-making perspective
on international affairs. See, for example, Tanaka, 1984; Hudson, 1983; and Thorson
and Sylvan, 1982. On the other hand, Al-type models have also been used to
understand DNA chains and the vision of frogs. Schrodt's use of pattern matching
in categorizing events is, in fact, based on the methodology used to analyze DNA
chains. (See Schrodt, 1984.)
7. Some would prefer the alternative term "algorithms." For a discussion of the
difference between the two terms, and the implications thereof, see Tamashiro, 1984.
8. A complete set of decision tree diagrams, along with the associated branch
logic, is available by request from the author. Request the Situational Predisposition
Decision Rule Index, and address inquiries to Professor V. Hudson, Department of
Political Science, Brigham Young University, Provo, Utah 84602.
9. The twenty-eight nations are: Canada, Cuba, Mexico, Venezuela, Chile, France,
Switzerland, Spain, West Germany, East Germany, Poland, Czechoslovakia, Yugoslavia,
the USSR, Norway, the Ivory Coast, Guinea, Ghana, Uganda, Kenya, Zambia, Tunisia,
Turkey, Egypt, Israel, the People's Republic of China, Japan, and the Philippines.
10. lt is important to recognize the distinction between the problem situation
that created an occasion for decision and the behavior (event) that resulted from
it. The coding procedure used to establish the situation was designed to allow us
to reconstruct the immediate external circumstances existing at the time the acting
government presumably recognized the problem (see Hudson, 1983). The resulting
event described the actor's subsequent behavior in response to the problem situation.
Coding of events was performed in the 1970s by one set of coders, while the roles
used to identify problem situations that precipitated the events were determined by
another set of coders a decade later.
11. This manual is available from the author: Professor V. Hudson, Department
of Political Science, Brigham Young University, Provo, Utah 84602.
12. lt should be noted that assessing the process validity (as versus the outcome
validity) of this test would not be meaningful, since the situational predisposition
model is not intended to simulate the processes by which the foreign policy choice
is made, only the outcome of that choice. Tangentially, it should be noted that
process validity does not necessarily entail outcome validity, or vice versa, in the
world of modeling. Because a model is a simplifying tool, there can be no absolute
achievement of one or the other and therefore no inevitable coupling of the two
concepts of validity. lt should be noted, however, that most modelers (especially in
AI) feel a certain (though unspecified) level of process validity is necessary for any
significant amount of outcome validity to occur. (This is reflected in the concept
of the "Turing test" of Al.)
13. Of course, here I beg the question of what is "reality," which involves
numerous intractable controversies, which I do not have the space to discuss in
detail. For this exercise, let us assume that it is possible to ascertain what, in reality,
nations and other international entities do or do not do in terms of foreign policy
behavior.
14. The issues that follow do not include mundane points like rendering one's
system outputs compatible with some extant archive of international events, the
quality of reporting of such events, "unobservable" behaviors (such as covert action)
not reported by such sources, and a whole host of others. It is assumed the
researcher has already braced himself or herself for such matters.
15. Here the 10 percent threshold is being used.
Bibliography
Achen, Christopher H. (1989) "When Is a State with Bureaucratic Politics Representable
as a Unitary Rational Actor?" paper presented at the annual meeting of the
International Studies Association, London, England, March 29-April 1.
Alker, Hayward, and C. Christensen (1972) "From Causal Modelling Artificial Intel-
ligence: The Evolution of a United Nations Peace-Keeping Simulation" in J.A.
Laponce and Paul Smoker (eds.), Experimentation and Simulation in Political
Science, Toronto, University of Toronto Press.
Azar, E. (1980) "The Codebook of the Conflict and Peace Data Bank" (Mimeo).
Banton, Michael (1965) Roles: An Introduction to the Study of Social Relations,
London, Taistock Publications.
Brady, L. (1978) "The Situation and Foreign Policy" in M. East et al. (eds.), Why
Nations Act, Beverly Hills, Sage Publications.
Bueno de Mesquita, B. (1981) The War Trap, New Haven, Yale University Press.
Callahan, P., L. Brady, and M. G. Hermann (eds.) (1982) Describing Foreign Policy
Behavior, Beverly Hills, Sage Publications.
Easton, David (1959) The Political System, New York, Alfred A. Knopf.
Etheredge, LloydS. (1987) "On Being More Rational than the Rationality Assumption:
Dramatic Requirements, Nuclear Deterrence, and the Agenda for Learning,"
unpublished paper.
Etheredge, Uoyd S. (1985) Can Governments Learn? American Foreign Policy and
Central American Revolutions, New York, Pergamon Press.
Goffman, Erving (1961) Encounters,lndianapolis, Bobbs-Merrill Educational Publishing.
Goodenough, Ward (1965) "Rethinking Status and Role" in M. Banton (ed.), The
Relevance of Models for Social Anthropology, New York, Praeger.
Gross, Neal, W. Masons, and A. McEarchen (1958) Explorations in Role Analysis,
New York, John Wiley and Sons.
Hermann, Charles F. (1983) "Superpower Involvement with Others: Alternative Role
Relationships," paper presented at the annual meeting of the American Political
Science Association, Chicago, Illinois.
Hermann, C.F., and M.G. Hermann (1979) "Summary of the External Components
in the CREON Project's Conceptual Framework for Explaining Foreign Policy,"
paper presented to the Northeastern Political Science Association.
Holsti, K.J. (1970) "National Role Conceptions in the Study of Foreign Policy"
International Studies Quarterly 14: 233-309.
Hudson, Valerie M. (1987) "Using a Rule-based Production System to Estimate Foreign
Policy Behavior: Conceptual Issues and Practical Concerns" in Stephen Cimbala
(ed.), Artificial Intelligence and National Security, Lexington, Lexington Press.
Hudson, Valerie M. (1983) "The External Predisposition Component of a Model of
Foreign Policy Behavior," Ph. D. dissertation, the Ohio State University.
Hudson, Valerie M., Charles F. Hermann, and Eric Singer (1985), "The Situational
Imperative," paper presented at the New Directions in the Comparative Study of
Foreign Policy Workshop, Columbus, Ohio, May.
International Monetary Fund/International Bank for Reconstruction and Development
(1964-1973) Direction of Trade.
Keesing's Contemporary Archives (1964-1973) Bristol: Keesing's Publications.
Kegley, C., Jr. (1980) "The Comparative Study of Foreign Policy: Paradigm Lost?"
Institute for International Studies Essays No. 10, Columbia, South Carolina,
University of South Carolina Press.
Lasswell, H. (1971) A Pre-View of Policy Sciences, New York, American Elsevier.
Linton, R. (1945) The Study of Man, New York, Appleton-Century-Crofts.
McGinnis, Michael, and J. T. Williams (1989) "Bayesian Rationality and Superpower
Rivalry: A Game Theoretic Model," paper presented at the annual meeting of
the Midwest Political Science Association, Chicago, Illinois, April 14-16.
Nadel, S. (1957) The Theory of Social Structure, Glencoe, Illinois, Free Press.
Rosenau, J. (1966) "Pretheories and Theories of Foreign Policy" in R. Barry Farrell
(ed.), Approaches to Comparative and International Politics, Evanston, Illinois,
Northwestern University Press.
Schank, R., and R. Abelson (1977) Scripts, Plans, Goals, and Understanding, Hillsdale,
New Jersey, Erlbaum.
Schrodt, Philip A. (1984) "Background to a Dynamic Model Using Artificial Intelligence
Techniques," paper presented at the U.S. and Swiss National Science Foundation
Conference on Dynamic Models of International Conflict, Boulder, Colorado, October
31-November 3.
Singer, Eric G. (1986) "Personality and Foreign Policy Behavior: A Study of Black
African Leadership," Ph.D. dissertation, the Ohio State University.
Singer, J.D. (1962) Deterrence, Arms Control, and Disarmament, Columbus, Ohio,
State University Press.
Small, M., and J.D. Singer (1973) "The Diplomatic Importance of States, 1816-
1970: An Extension and Refinement of the Indicator" World Politics 25: 577-
599.
Steinbrunner, J. (1974) The Cybernetic Theory of Decision: New Dimensions of
Political Analysis, Princeton, Princeton University Press.
Tamashiro, Howard (1984) "Algorithms, Heuristics, and the Artificial Intelligence
Modelling of Strategic Statecraft," in D.A. Sylvan and S. Chan (eds.), Foreign
Policy Decision-Making: Perception, Cognition, and Artificial Intelligence, Praeger,
New York.
Tanaka, Akihiko (1984) "China, China Watching, and CHINA-WATCHER" in D.A.
Sylvan and S. Chan (eds.), Foreign Policy Decision Making: Perception, Cognition,
and Artificial Intelligence, Praeger, New York.
Taylor, C., and M. Hudson (1972) World Handbook of Political and Social Indicators,
New Haven, Yale University Press.
Thorson, S.J. (1984) "lntentionallnferencing in Foreign Policy: An Artificial Intelligence
Approach" in D.A. Sylvan and S. Chan (eds.), Foreign Policy Decision-Making:
Perception, Cognition, and Artificial Intelligence, Praeger, New York.
Thorson, S., and D. Sylvan (1982) "Counterfactuals and the Cuban Missile Crisis,"
International Studies Quarterly 26: 539-571.
Tukey, J. (1977) Exploratory Data Analysis, New York, Addison-Wesley.
U.N. Statistical Yearbook (1960-1973) New York: Publishing Services of the United
Nations.
United States Arms Control and Disarmament Agency (1973) World Military Ex-
penditures and Arms Trade, Washington, D.C., Government Printing Agency.
Walker, S. (ed.) (1987) Role Theory and Foreign Policy Analysis, Durham, N.C., Duke
University Press.
9
UNCLESAM: The Application of a
Rule-Based Model of U.S. Foreign Policy Making
Brian L. Job and Douglas Johnson
Among the several modes of formal, computational modeling involving

artificial intelligence methods and technologies, rule-based systems continue
to occupy an important place. Many of the earlier efforts to model the
behaviors of decision-makers and institutions in foreign policy and inter-
national environments were rule-based systems (e.g., Alker and Christensen,
1972; Thorson and Sylvan, 1982). Rule-based models have come under
critical scrutiny recently, largely on the basis of their limitations to represent
behaviors in complex and unspecific domains (see Mallery, 1987). However,
both for logical and practical reasons, rule-based architectures continue to
be prevalent. In many circumstances the behavior of individuals or agencies
is, at least in part or at certain times, appropriately characterized as rule-
based. Alternately, the limitations of available resources and/ or available
technologies necessitate the resort to the use of rules, and rule-based logics,
in modeling efforts that would otherwise seek to utilize alternative architectures
(see Mefford, this volume).
In this chapter, we report on our initial efforts to develop and apply
models of U.S. foreign policy toward states in Central America and the
Caribbean. 1 Here, we adopt a rule-based modeling approach and seek to
demonstrate that a computational model with a basic control structure
operating with a small set of rules can accurately replicate a history of U.S.
foreign policy outputs. The specific context to which our model, labeled
UNCLESAM, is applied is a case study of U.S. policy and actions toward
Work reported upon here was supported through research grants from the University of
Minnesota (Department of Political Science and College of Liberal Arts). The authors thank
James Mahoney, Eric Selbin, and Valerie Hudson for their comments on an earlier draft.
Particular acknowledgment is due Eric Selbin for his assistance concerning materials on Central
America and the Caribbean. Any errors or omissions remain our responsibility.
221
222 Brian L. Job and Douglas Johnson
the Dominican Republic from 1959 to 1965. In that we are also especially
concerned that our modeling efforts facilitate exploration of issues of
theoretical importance to the study of foreign policy, UNCLESAM is designed
with the architecture, i.e., the structural and operational properties, of a
modified cybernetic decision-maker. Thus, application of this model to the
Dominican Republic case study provides opportunity to assess, in a systematic
fashion, the extent to which U.S. policy outputs could be characterized as
if they were the product of a cybernetic actor. Our overall research strategy
for the utilization of formal, computational models entails a logic of building
from simple, noncomplex formulations and models, taking note of where
such models succeed and fail, and going on to models with more sophisticated
architectures and broader domains of application.
This chapter is organized as follows: The first section describes the central
characteristics of U.S. foreign policy decision making in the Central American/
Caribbean region as noted by key scholars in this field. Important parallels
between this substantive description and the analytical description of cy-
bernetic decision making, set forth by Steinbruner (1976) and others, are
noted. In light of particular features of the U.S./Central American context,
however, modifications to the standard cybernetic decision-maker are required.
The second section, in turn, goes on to indicate how the features of this
modified cybernetic decision-maker may be represented in the format of a
rule-based computational model, in general terms, and in specific terms by
the UNCLESAM model. The application of UNCLESAM to the Dominican
Republic case is described in the third section. In the concluding section,
attention is turned to assessment both of the relative success of UNCLESAM
in replicating the specific sequence of U.S. behaviors re the Dominican
Republic and of the larger issues of the plausibility and external validity
that UNCLESAM possesses as a model of U.S. foreign policy decision
making.
Designing a Model of U.S.-Central American

State Interaction
U.S. Foreign Policy Decision Making Concerning Central America

Throughout the past century Washington has come to regard the Caribbean
as an "American lake" and Central America as unequivocally within the
United States' most "legitimate" and immediate sphere of influence. The
record of U.S. policy toward this region is a paradoxical juxtaposition of
inattention and ignorance toward these states and their circumstances, on
the one hand, combined with pervasive economic penetration and persistent
gunboat diplomacy, on the other. (Thus, for example, Lowenthal, 1982, and
Wiarda, 1984, emphasize the former factors, whereas LaFeber, 1983, con-
A Rule-Based Model of a.s. Foreign Policy Making 223
centrates upon U.S. economic domination.) On at least forty-five occasions

between 1948 and 1976 the United States employed its military forces in
Central America to advance its policies (Job and Ostrom, 1986). Of the
five most significant military undertakings within this set, the invasion of
the Dominican Republic in 1965 was the largest operation, involving thousands
of troops and lasting over a year (Lowenthal, 1972, and Slater, 1978).
Within this arena of great-power domination, U.S. policy-makers and
policy making took on several distinct characteristics: First, although one
or more states besides the United States might be involved, relationships
were effectively unilateral. This absence of bilateral character to U.S. relations
was due to the gross power differentials involved and to the U.S. failure to
consider these states as having their own legitimate interests and goals.
Second, deeply rooted, simplistic stereotypes held sway over U.S. thinking
and action. Washington came to view all Central American/Caribbean
countries according to a single, basic stereotype. The stereotypical Central
American/Caribbean state involved a passive population, susceptible to
agitation by internal and external opposition forces. Popular, indigenous,
democratic opposition to the regime in power was regarded as impossible.
External, left-wing (read "communist") forces were always suspect, leading
successive U.S. administrations to support and bolster friendly regimes that
offered the prospect of stable relations, albeit at the expense of the population's
economic and social well-being and political freedom. 2
Third, U.S. policy responses became routinized and cued to reactions to
relatively unsophisticated indicators of conditions within Central American/
Caribbean states. Concerning the formation and execution of U.S. policy,
within the larger context described above, Lars Schoultz (1987) in his recent
book National Security and U.S. Policy Toward Latin America argues that
there are but "three crucial values that orient the policy making process:
simplicity, stability, and security" (Schoultz, 1987: 29). 3 The latter two,
stability and security, i.e., U.S. security, according to Schoultz are inextricably
linked in the minds of decision-makers.
U.S. policy . . . is largely determined by the manner in which one fact,

instability, is simplified when it is perceived by U.S. policy makers.
[A)Il U.S. policy makers feared instability, regardless of its cause, because of
its potential negative consequences for U.S. security (Schoultz, 1987: 11, 24).
(Jn)stability, thus, serves as the crucial and summary indicator for U.S.
oflicials. 4 Outbreaks and perceived increasing levels of instability are regarded
as involving increasing threats to U.S. interests in the Central American
state and by extension to U.S. security itself. Decisions and actions are, in
turn, taken to counter the forces/agents perceived to be the causes of
instability with the goal of restoring the status quo. In some instances,
because new leaders in a regime are seen to be the problem, or because
opposition forces threaten the regime in power, the United States' attitude
or posture toward the country will change. Thus, a country undergoing civil
disorder (instability) with the prospect of a "left-leaning" government taking
power is apt to experience the effects of a dramatic reversal in U.S. favor
from positive aid and support to negative sanctions, perhaps even overt
intervention. In contrast, a friendly regime beset with instability caused by
popular discontent will likely find its support from Washington substantially
increased, at least until or unless the United States loses patience with the
leadership's inefficiency.
Simplicity figures in the policy process in two ways: as a descriptor of
the manner in which information is monitored, channeled, and filtered up
through the levels of bureaucracy; and as a value, in the sense that decision-
makers, especially at higher levels, desire to be presented with uncomplex
characterizations of events and problems in Central American states and
with straightforward choices from "menus" of available, and proven, policy
alternatives. Schoultz (1987: 26) offers the following analogy to describe the
workings of the process:
Senior policy makers assign lower-ranking officials to watch over U.S. interests
in Latin America. Under normal circumstances these [persons] do their job in
much the same way that art museum guards watch over their paintings. They
are essentially guardians. . . . Every so often, however, instability flares.
Someone (or some group) starts running around the galleries of the Latin
American museum . . . generally disturbing the desired atmosphere of calm.
When this occurs, the guards sound the alarm, [and] senior officials come
running.
Finally, U.S. policy responses will differ little from country to country.
Since a basic stereotypical interpretation is applied to all, the overriding
assumption is that what works in one situation, with one leader, against
one leftist opposition movement, etc., will work in any other. The United
States has at its disposal a full array of policy responses, ranging from
diplomatic persuasion to full-scale military deployment. High-level officials,
therefore, want to be presented with simplified statements of the situation
at hand, (that is, extent of threat to U.S. interests derived from an assessment
of instability), along with recommended policy options that follow the formulaic
patterns of previous U.S. behavior toward Central American states (see
Etheredge, 1985).
A Rule-Based Model of U.S. Foreign Policy Making 225
A Modified Cybernetic Model of

U.S. Foreign Policy Decision Making
We viewed the substantive description summarized above as yielding the
following salient assumptions as the bases for the architecture of our model.
First, the United States can be represented as a unitary actor vis a vis
its foreign policy behavior toward Central American states. Representation
of separate U.S. agencies, bureaucracies, or individuals did not appear
necessary or warranted in the first phase of our modeling efforts.
Second, a single-actor model is sufficient. Albeit that there were two
states involved, i.e., the United States and the Dominican Republic or some
other state(s), the United States simply functioned on the basis of its own
perceptions, interests, and goals without having to consider its relationship
in any contingent bilateral fashion. Furthermore, U.S. decisions were essentially
a series of reactive responses. In the absence of activity within the Central
American state, the United States took no action or made no policy changes.
In fact, long periods in which no obvious attention is being paid and there
are no interruptions in U.S. diplomatic, military, and economic relations,
positive or negative, are common. Only when events occurred that triggered
concerns about stability or security does the United States actively monitor
the situation and take a policy action.
Third, the process of decision making can be represented by a simple
architecture involving three steps or phases: information monitoring and
input, assessment, and decision-leading to policy action. As per Schoultz's
analogy, the monitoring and input phase corresponds to the activities of
lower-level U.S. officials who review ongoing events in Central American
countries specifically attentive to anything that might impact upon levels
of (in)stability. The assessment phase refers to the interpretation of these
events regarding levels and changes in stability that are reported to higher-
level decision-makers. Finally, the decision phase corresponds to the decisions
taken by higher-level decision-makers who, having established a posture
toward the regime in question, choose a policy response from the inventory
of available U.S. actions.
The student of foreign policy decision making will recognize key parallels
between this process and the prototypic cybernetic decision-maker and
decision making featured in the foreign policy literature. Note, for instance,
Steinbruner's (1976: 51) description of cybernetic decision making as "or-
ganized around notions of short-cycle information feedback and the elimination
of uncertainty." Further note his depiction of the cybernetic decision-maker,
following upon Ashby's characterization of homeostatic behavior, as func-
tioning to achieve ultrastability through the monitoring of critical variables
and the taking of action only as the values of these variable move outside
self-defined tolerable ranges. As to its responses, the cybernetic decision-
maker "carries a repertory of behavior patterns each of which operates in
a characteristic way upon receipt of perceptual input" (Steinbruner, 1976:
53-54).
In that these statements capture the basic features of U.S. foreign policy
behavior presented earlier, our model is designed accordingly as a cybernetic
actor. However, two modifications to the standard cybernetic prototype are
required. Observation of U.S.-Central American interaction indicated that
some form of memory function was required in our model, not a particularly
sophisticated memory but rather one that remembers the status of U.S.
relations with the other state and the immediately prior set of U.S. actions
toward the state. 5
A second modification was necessary in order to capture the context-
dependent character of U.S. action. In each specific situation of time and
place, U.S. response to the country in question would be contingent upon
its diplomatic stance toward the regime in power. (We use the term "posture"
to refer to the general attitude of the United States toward another country
and its government.) For example, in the context of a negative posture
toward a country, the United States might impose sanctions upon or intervene
in the affairs of a Central American/Caribbean state, while the same
circumstances may produce extensions of aid and support to a regime the
United States regarded positively. Representation of this context-dependent
character of decision making necessitated delineation of a model with a
two-stage decision process. Before taking action to respond to the situations
that triggered concern over instability (step two), the modeled decision-
maker first checks on the current U.S. posture toward the Central American
state (step one). With the addition of these features, we came to characterize
our model as one representing "modified, two-staged cybernetic decision
making."
UNCLESAM as a Computer Program

The formal, computational model we designed representing U.S. foreign
policy action as a modified cybernetic decision-maker was called UNCLESAM. 6
With UNCLESAM, output decisions are produced in an ordered framework
or hierarchy involving the repeated application of rules to a knowledge base
that changes over a sequence of events. UNCLESAM, therefore, could be
described as a formal system, implemented as a computer program, consisting
of symbolic representations of historical events, rules, and metarules that
enable and disable other rules and a control structure that determines the
order of applying metarules to rules and rules to the knowledge base. After
providing a schematic description of the architecture of UNCLESAM, in the
third section the specifics of its application concerning the United States
and the Dominican Republic are detailed.
The Architecture of UNCLESAM

Figure 9.1 indicates the main modules of UNCLESAM's decision archi-
tecture. To reiterate, UNCLESAM's design centers upon the representation
of a decision-maker with a knowledge base of facts and assessments of
critical values, who, upon receiving a temporally ordered sequence of events
data, takes decisions through the systematic application of predetermined
rules. A "control structure" functions to keep the model operating, i.e., to
maintain and order the sequencing of the stages of event input, knowledge
representation, and application of rules, as well as the various process steps
within these stages. Understanding the way UNCLESAM functions depends
on appreciating the definition and role of its knowledge base and of its
rules-of which there are four types: time-dependent rules, assessment rules,
metarules, and decision rules (per Figure 9.1).
In order for UNCLESAM to "make decisions," it must first be presented
with timely information concerning the situation at hand. However, such
information must be translated into a format that UNCLESAM can "un-
derstand." Thus, as events take place, they are represented as semantic
networks in a specialized language, PAL (political action language) developed
for this purpose and placed in UNCLESAM's knowledge base. The political
situation at any point in historical time is the environment created up to
that juncture by the PAL representations of the networks linking events,
actors, attributes, and actions.
At this juncture, UNCLESAM applies its rules to the knowledge base,
starting with the application of assessment rules. These are rules that
calculate if events or conditions have altered sufficiently to produce new
values on the indicators critical to the decision-maker. In UNCLESAM, the
single critical indicator, of course, is perceived stability within the Central
American state.
Changes in values of this critical indicator, in turn, will trigger the
application of the decision rules in the next step in UNCLESAM. However,
decision rules are context dependent, in that first metarules must be applied
to determine which set of decision rules are relevant in light of the U.S.
posture toward the Central American/Caribbean state at that moment. A
separate set of decision rules is defined for each different U.S. "posture
state." Once UNCLESAM is located within the appropriate context, its
process continues by instantiating specific decision rules. When the ante-
cedents of a decision rule successfully match facts in the knowledge base,
the rule "fires" and the actions/decisions specified are taken. These actions/
decisions, in turn, are events, information about which is added to the
Figure 9.1: The Architecture of the UNCLESAM Model
~
CXl
( CONTROL J
_,_..
_,----'-' --------\;_,--------------------
', .------'=:._ _ _ _ _--....
_,_.._.. \. CONTEXT DEPENDENT

_,/ _..- "-J
r------'~----.... ~,---~-------
ASSESSMENT RULES
NEXT KNOWLEDGE BASE

EVENT
CONTEXT SPECIFIC
NEW I ASSESS- ACTION
f I ONGOING DECISION RULES
FACTS FACTS MENTS EVENT
HISTORICAL
EVENT
STREAM
knowledge base. When the control structure causes this to happen, it invokes
the decision algorithms in UNCLESAM recursively on the now modified
knowledge base. The decision-maker in this sense recognizes its own actions.
Unmentioned thus far have been the time-dependent rules noted in Figure
9.1. These rules appear at first glance to be rather curious in that they
suggest that responses are made in the absence of any stimuli. However,
they are very necessary for UNCLESAM's maintenance of an ongoing,
dynamic knowledge base. These rules, which must be explicitly formulated
in a computer model, represent what is the normal, but implicit, behavior
of human decision-makers in circumstances when "no events" take place.
During such periods, what is usually happening is the incremental reas-
sessment of critical indicators in reaction to the gradual damping down of
levels of prior high activity. (For example, it is common to find no new
civil disorders occur following application of riot control measures by the
police. Accordingly, after some point in time, values of stability or order
must be reassessed.) UNCLESAM, if not provided with rules to accomplish
stability reassessment in such circumstances, will continue, in cybernetic
fashion, to maintain its indicator values and to respond accordingly to later
inputs by further escalating its assessed instability level. Thus, "nonevents"
are defined as periods of time when nothing is reported, which is itself
significant for the situation being monitored. The relevant fact in this case
is time itself, specifically, the length of time that elapses. To handle these
circumstances, UNCLESAM incorporates rules that assess the effects of the
passage of time. Mechanically what happens is a preprocessing of the
knowledge base as a new historical event fact is about to be added. If no
events have occurred for a specified interval, updated (usually decremented)
values of conditions are applied to those that UNCLESAM would otherwise
presume to be in place.
The Sequence of Operations in UNCLESAM

Thus, to review: Whenever a new event "occurs" and is to be added to
UNCLESAM's knowledge base:
1. The events are represented in the semantic networks of the political

action language, "understood" by UNCLESAM.
2. The time-dependent, nonevent rules are applied as necessary.
3. The knowledge base is updated to reflect assumptions about the new
facts. Consistency checks are carried out.
4. The assessment rules are applied to the knowledge base. This may
result in new values of critical variables-these facts also to be added
to the knowledge base.
5. The context-dependent metarules are applied to determine which

decision rule(s) to apply.
6. The relevant decision rules are applied to the knowledge base. This
may trigger the production of decisions/actions.
7. These new event facts are added to the knowledge base, which
recursively invokes the process to commence again.
Representation
Further comment is required concerning the manner in which event

information is represented in UNCLESAM using the medium of the designed
PAL. The raw event data from which a knowledge representation is constructed
consists of text descriptions of relevant events in the domain under con-
sideration. While this textual description may be exceedingly complex, the
cybernetic paradigm suggests that a simplified conceptual view is significant
and sufficient to decision making. In line with this, we developed a narrowly
defined event representation language system composed of five semantic
primitives: events, actors, actions, relations, and attributes. For each class
of primitives, we defined a small set of mutually exclusive categories. These
categories were generated in an ad hoc fashion as events were interpreted.
The level of representation used, however, remains purposely coarse-grained.
Actions are the central primitives of the PAL system. Each action has
a slot for an actor, an object (another actor or action), and may also have
an intensity slot. 7 Actions are organized according to a classification scheme
that represents: (1) knowledge of the effects actions have on the objects
of actions-usually other ongoing actions but also relationships and attributes,
and (2) the duration of various actions. Ongoing actions are distinguished
from one-shot actions. Interpretation of an event that describes the ending
of a condition or event results in the removal of the initial condition or
event fact from the knowledge base. In the jargon of artificial intelligence,
functions that accomplish these tasks of ensuring consistent and updated
representation of facts throughout the model's knowledge base are known
as "truth maintenance" functions.
Relations primitives represent static relationships or connections between
actors, the most common relations primitives involved in UNCLESAM being
the various political relationships between actors, such as leader (of the
regime), president (of the state), and diplomatic relations. Some relations
can also have an intensity. Thus, the most important relation category,
posture, varies along five levels of intensity. Attribute primitives represent
qualities or properties intrinsic to the actors. But only one category of
attribute, ideology, is important in UNCLESAM. 8
Because actors are defined as the agents of purposeful actions, the
representation of actors in PAL is linked to the kinds of actions and relations
Figure 9.2: Three Examples of the PAL Semantic Network Representation of Events
in the UNCLESAM Model
OR MHftary attacks (moderately) Rebels
Balaguer
Event-228 -Threaten<
OR-Government
Escalate-Action-Suppress - - - -
---Opposftion
Balaguer threatens to have the government increase fts suppression of the opposftion.
Event-41- Change-Attribute (-1) < Trujillo
Ideology
TrujUio shifts (his ideology) to the left
they are arguments of. Individuals, groups, churches, governments, states,

and groups of states are all represented as instances of actors. In the
UNCLESAM model there is one specialized kind of actor, the United States,
which has decision-making capabilities.
Figure 9.2 presents examples of the manner in which each raw event
was described as a single assertion represented by a semantic network. As
this was a laborious manual task, software tools were developed to facilitate
constructing and editing the historical sequence of represented events that
in turn was to be processed by the decision-making rules in UNCLESAM.
Application of the Rule-Based UNCLESAM Model

to a Case History of U.S. Foreign Policy
The Dominican Republic, 1959-1965

Relations between the U.S. and the Dominican Republic from 1959 to
1965 present an ideal case for our modeling venture. The United States had
a long-established tradition of managing affairs in this country, including an
extended occupation of the country by U.S. Marines early in the twentieth
century, and the imposition of Trujillo, Sr., as a leader to be sympathetic
to United States interests. Within the scope of the six years chosen for our
case study, the United States at one time or another employed virtua.lly
every available instrument of policy statecraft. Thus, on the diplomatic
dimension the United States moved from having full diplomatic relations to
having no diplomatic relations, with varying stances in between. On the
economic dimension, the United States at different times encouraged trade,
threatened to cut imports, did cut imports, and imposed other economic
sanctions. With regard to the use of military force, U.S. actions ranged
from no action, through stages of readiness, alert, deployment, demonstra-
tions, to a complete invasion by the marines. From the time of the
assassination of Trujillo Sr. in 1961 to this invasion in 1965, there were six
different governments in the Dominican Republic. The U.S. posture toward
each one varied across the spectrum from warm support to hostile opposition.
A capsule history of the six-year period includes the following key
happenings (Lowenthal, 1972; Slater, 1978): By 1959 U.S. patience with the
Trujillo regime, which it had strongly supported for years, was wearing thin.
Washington was nervous with its partnership with the repressive dictator,
in light of developments in Cuba, but it could identify no viable alternative
leadership. Trujillo's increasingly repressive behavior, obvious popular dis-
content in the Dominican Republic, and the condemnation of the Trujillo
regime by the OAS led the United States to effectively break diplomatic
relations, end its aid, and apply economic penalties to the Dominican Republic
by August 1960. In May of 1961, Trujillo, Sr., already having resigned the
presidency under pressure, was assassinated. His successors, including his
son, clamped down even harder on the population. Opposition and government
forces continued to struggle, however, leading the United States in November
1961 to undertake a significant and bellicose display of force off the coast
of the Dominican Republic in an effort to thwart Trujillo, Jr., and his allies.
This gunboat diplomacy was successful, in that Trujillo and his cohorts
resigned and fled. A year later, Juan Bosch was the victor in democratic
elections and came into office with strong U.S. support. Within three months,
U.S. aid was withdrawn because of his "revolutionary" reforms. In September
1963 Bosch was overthrown in a coup. The Kennedy administration's
ambivalence toward this new regime (it withheld U.S. recognition) was soon
replaced by Johnson's warm embrace. Thus, 1964 is notable for the massive
amounts of U.S. assistance provided the Dominican Republic.
Despite these efforts, domestic political conditions remained uncertain
and quietly, at first, and then not so quietly deteriorated into 1965. In April,
military factions led a revolt that sought to restore constitutional government.
The regime ineffectually attempted to counter these efforts. The United
States, initially supportive of the regime, became dismayed at both its
failures and the prospect of leftist elements gaining power in the country.
In the face of growing civil disorder, the United States with almost no
A Rule-Based Model of O.S. Foreign Policy Making 233
preliminary actions took matters into its own hands, mounting a full-scale
invasion and takeover of the country.
An examination of this record, in light of our intention to design a model
to generate U.S. policy responses, led us to the following conclusions: First
of all, priority would be given to modeling U.S. use of force activities. Not
only were these behaviors the most dramatic and salient on the part of
Washington, they also appeared to be most directly related to assessments
of perceived instability within the Dominican Republic.
Second, shifts in U.S. posture, i.e., shifts along a continuum of positive
to negative disposition toward the current Dominican Republic regime, were
clearly of central importance. All decisions concerning economic relationships,
including exports, imports, and aid, appeared keyed to, and apparently timed
with, U.S. posture changes. However, while their significance for resulting
U.S. policy behavior was obvious, the manner in which U.S. posture shifts
vis a vis Dominican Republic regimes related to contemporaneous events
in the republic was not apparent, even to a careful reader of the historical
record. Thus, we were presented with a dilemma in our modeling enterprise-
we could identify instances of U.S. posture shifts and the resultant policy
actions that appeared to follow these shifts, but we could not ascertain the
nature of the decision process that led to the shifts themselves. We decided,
therefore, to consider the shifting of postures as outside, or exogenous, to
the cybernetic decision process represented in UNCLESAM. UNCLESAM,
then, would take decisions in light of current posture status but would
receive information as to posture shifts via its input event stream, as opposed
to deciding upon these posture shifts itself. This important limitation, along
with its revealed consequences in the application of UNCLESAM, is further
discussed in the last section.
UNCLESAM: Application and Results

Construction of Event Data Information Input to UNCLESAM. UNCLESAM
produces decisions in response to its "perception" of, and assessment of
the consequences of, information it receives about ongoing events. For our
purposes, the information stream relevant to UNCLESAM was taken to be
represented adequately by the series of events concerning U.S.-Dominican
Republic interaction reported in the New York Times from January 1956
until the U.S. invasion in April 1965. Care was taken to ensure that
contemporaneous U.S. government understandings and perceptions of what
was going on in the Dominican Republic, rather than subsequently corrected
or reinterpreted versions, were recorded. Narrative sources, including Low-
enthal's (1972), Slater's (1978), and Martin's (1966) accounts of these times,
were consulted as well. 9 During this period, the New York Times carried a
total of 459 reports on the Dominican Republic. From these reports, 137
Figure 9.3: UNCLESAM Representation of the Dominican Republic
A. ACTIONS/ATTRIBUTES OF THE DR ENVIRONMENT

1. President
2. Vice-President
3. DRG Suppression of Opposition

4. Opposition to DRG
5. Stability of DR. Government
B. ACTIONS/ATTRIBUTES OF THE US. REGARDING THE DR

6. Posture Towards the DRG
7. Diplomatic Relations with the DRG
8. Exports to the Dominion Republic
9. Imports from the Dominion Republic
10. Economic Aid to the Dominion Republic
11. Military Aid to the Dominion Republic
12. US Use of Force towards the DRG
NOTE:
Levels of Actions/Attributes 1.-4. above determined by coding of Events

in the OR
Level of Stability of DR Government determined by assessment
procedures in UNCLESAM
Levels of 7.-11. above determined by UNCLESAM in response to
Posture Shifts.
Levels of US Use of Force determined by UNCLESAM in response to
Level of Stability Assessment, conditioned by Posture Level
separate events, including changes of conditions, happenings in the Dominican

Republic, as well as decisions taken by the United States, were identified.
These were, in turn, translated into PAL semantic network statements and
input to UNCLESAM. These statements were in the form of those illustrated
in Figure 9.2.
Figure 9.3 indicates the format of UNCLESAM's representation of the
Dominican Republic within the model's knowledge base. At any point in
time, the current state of the situation of conditions in the Dominican
Republic and U.S.-DR relations is described by the set of twelve attribute/
action values recorded in UNCLESAM. For attributes such as office holding,
e.g., president, the relevant value was an actor's name. Conditions such as
level of opposition, level of government suppression of opposition, and levels
A Rule-Based Model of 0.5. Foreign Policy Making 235
of exports/imports were recorded on ordinal, seven-point scales. Stability

of the Dominican Republic government, the critical variable for our model,
was also assessed according to a seven-point scale. 10 U.S. posture toward
the Dominican Republic was defined over a continuum with five ordinally
located positions, ranging from an unmitigated positive disposition to a
completely negative stance toward the current regime. The condition of the
application of U.S. use of force toward the Dominican Republic was defined
as ranging over seven possible stances or deployments. 11 Recall that UN-
CLESAM had to be "initialized," i.e., the model had to be provided an initial
state of knowledge at the beginning of its run. Thus, for each of the
attributes/actions, UNCLESAM was set up with a value representative of
conditions of January 1, 1956. From that point on, changes in these values
occur either as a result of event information inputs or as a result of actions
taken by UNCLESAM as the model's decision rules were triggered.
Operation of UNCLESAM: Application of Rules. As a rule-based model,
UNCLESAM operates to produce assessments of the critical variable (stability)
and to produce decisions/actions of U.S. policy through the application of
the four types of rules.
The stability assessment rules, thirteen in all, are cued to particular event-
type occurrences, as described in the PAL statements. Thus, for example,
occurrence of an "attack against the government" of a particular intensity
level will result in the application of a rule leading to an increment in the
assessed level of instability in the Dominican Republic. In complementary
fashion, ending of civil unrest, disorders, government repression, etc., will
result in the decrementing of the perceived level of stability. Note that the
stability assessment rules are designed to produce results that reflect
decreasing marginal effects of continued unrest on perceived stability levels.
The occurrence of a riot or minor attack against the government will have
a relatively large incremental impact on assessed instability, provided that
stability levels are low (e.g., 3 on the scale) when these events happen. If
already at a high level of instability, repeated occurrences of such events
will not result in the instability assessment being raised. Only correspondingly
more intense activity can lead to yet higher levels (e.g., 6 or 7 on the scale).
If the frequency and intensity of events is not sustained, UNCLESAM will
decrement its assessed instability level, possibly in response to the triggering
of its time-dependent rules.
The U.S. attitude toward the Dominican Republic regime was represented
as a posture, of which there were five defined types, although only four
were demonstrated in the application of our model. These ranged from
Posture 1 as a most friendly I supportive stance to Posture 4 representing
a negative/ sanctioning stance. As discussed earlier, a shifting of posture
level results in the determination of the levels of diplomatic relations, exports,
imports, and aid between the United States and the Dominican Republic.
Thus, with each posture shift, a set of rules are triggered that produce
these level adjustments. To illustrate with an example of such a rule:
IF U.S. Posture to the DR Govt (<4), changed to 4,
THEN Decrement Diplomatic Relations level ( -1 ),

Decrement U.S. Imports Level (- 2), and
Decrement U.S. Exports Level (- 2).
UNCLESAM's "decisions," i.e., U.S. policy actions regarding the Dominican

Republic, resulted from the application of the decision rules in response to
altered assessments of the stability level in that country. These decision
rules are context dependent; thus, different rules are applied depending upon
the current U.S. Posture condition. A total of fifteen decision rules were
required, each of the following form:
IF U.S. Posture to the DR Govt 4,
and Stability Level > = 5,

and Stability Level change >0,
THEN Increment (+ 1) U.S. Use of Force Level.
Two time-dependent rules were utilized. Their operation was triggered by

the absence, for specified time periods, of· reported events concerning the
nature of conditions in the Dominican Republic. Thus, for instance, in the
circumstance of no reports related to instability over a period of two weeks,
given a stability assessment at level 4 (because of previously acted-upon
event facts), the stability assessment would be reduced to level 3. However,
during the application of UNCLESAM, a related problem arose that such
time-dependent rules could not address. This concerned "unreported" events,
that is, events the New York Times did not report on the date of their
occurrence, but events that could be inferred to have taken place, given
reporting of subsequent events. For UNCLESAM, "unreported" events were
a particular problem concerning the reduction and withdrawal of U.S. force
deployments toward the Dominican Republic. Initiation of use of force levels
were followed very attentively by the U.S. news media; withdrawal of force
units was not. This was not a trivial problem for UNCLESAM, a model that
makes decisions to increment or decrement policy responses, contingent
upon what it "knows" to be the current level of such activity. In a notable
example, we found that UNCLESAM, understanding through its knowledge
base that the United States was deployed at a certain level, "decided" that
the United States should invade the Dominican Republic. An examination
revealed that the model's knowledge base had continued to record the level
of U.S. use of force deployed months earlier but long since withdrawn. In
A Rule-Based Model of il.S. Foreign Policy Making 237
Figure 9.4: UNCLESAM Representation of Dominican Republic Events/Actions of June

1961
US posture to DRG = 4, Stability of DRG =3
6/1/61 Trujillo Assassinated

••• US Assessment of Stability, Level = 4
••• US Use of Force, Level= 3
6/2/61 Trujillo Jr. Attacks {level= 5) DRG
•• US Assessment of Stability, Level = 5
••• US Use of Force, Level = 4
6/3/61 Trujillo Jr. Stops Attacks on DRG

•• US Assessment of Stability, Level= 4
6/3/61 Trujillo Jr. Attacks {level= 5) Opposition
•• US Assessment of Stability, Level = 5
••• US Use of Force, Level = 5
6/9/61 Trujillo Jr. Changes Attacks {level= 3) on Opposition
6/15/61 DRG Changes Suppression of Opposition {level = 2)
+++ US Chanies Posture to DRG, Level =3

••• US ssessment of Stability, Level = 3 (2 weeks)
NOTE:
••• indicates response produced by UNCLESAM model
+++ indicates posture change, an mput to UNCLESAM
five instances we were required to directly input event information to

UNCLESAM to compensate for unreported changes in U.S. actions. 12
UNCLESAM: The Success of the Model

There are a variety of issues to consider here concerning UNCLESAM.
For any rule-based production system, replicability is a critical criterion.
That is, to what extent is the model capable of producing, through the
application and reapplication of its rules, behavior patterns that duplicate
the observed sequence of U.S. policy actions in the case(s) under investigation?
However, there are no uniformly accepted standards or "tests" that can be
applied to provide a simple answer to this question (Mallery, 1987). If we
focus upon the particular outcomes of most interest to us, namely, U.S.
Figure 9.5: UNCLESAM Representation of Dominican Republic Events/Actions of April

1965
US Posture to DRG =2, Stability of DRG =3
4/24/65 PRO Attacks (level = 3) DRG

4/24/65 DRG Raises Suppression of Opposition (Level= 6)

4/25/65 OR Army Takeover (Coup) DRG
+++4/25/65 US Changes Posture to DRG, Level = 3

••• US Changes Dip Rei to DRG, Level = 2
4/25/65 Rebels(DR Airforce) Attacks DRG (OR Army), Level = 3
4/25/65 ORB (OR Army) Attacks REbels (OR Airforce), Level = 4

*** US Use of Force, Level = 7
NOTE:
••• indicates response produced by UNCLESAM model
+++ indicates posture change, an mput to UNCLESAM
decisions regarding the use of force toward the Dominican Republic, and
we examine UNCLESAM's capacity to replicate the four major escalatory
sequences of U.S. action during 1959-1965, we find that UNCLESAM
performed quite "satisfactorily." Thus for the crises of 1961 and of 1965,
and in the two other situations involving U.S. uses of force, UNCLESAM
accurately reproduced a record that mirrored U.S. actions in these circum-
stances. (Figures 9.4 and 9.5 are provided as an illustration of UNCLESAM's
results in 1961 and 1965.) Further concerning use of force policy sequences,
UNCLESAM produced one "false positive" output, i.e., a decision to deploy
force when, in fact, the United States did not take such an action. There
were no "false negative" results, i.e., instances of U.S. use of force decisions
that UNCLESAM failed to produce. Thus, our model, applied over a six-
year period, was capable of assessing conditions in the Dominican Republic,
identifying the times when the United States would undertake uses of force,
and tracking the pattern of these actions. 13
But replication of behavioral output, although important, is, by itself, of

limited interest beyond a certain point. Of larger significance is determining
if the structure and logic of the model's architecture represents, in a
"theoretically meaningful" way, the process of assessment, decision, and
action of U.S. foreign policy behavior. Plausibility and theoretical relevance,
thus, combine as a second set of criteria for assessing UNCLESAM. Note
that the phrase "theoretically meaningful" is for us a quality defined by the
nature and purposes of our investigation. Thus, we turn to consideration of
how UNCLESAM, as representing a prototypic, modified, cybernetic decision-
maker, functioned in this application. In this regard, we believe that we
have expressed and applied, in a formal fashion, a model that demonstrates
the plausibility of the substantive arguments put forth by Schoultz (1987)
concerning U.S. foreign policy toward Central American and Caribbean states.
UNCLESAM, with its limited interpretative capacity, limited set of rules,
attention to a single indicator (instability), and consistent reproduction of
patterned responses, showed that an actor operating according to this limited
decision-making structure could indeed produce outputs similar to those of
the United States in an historically important case. The cybernetic prototype
does appear to have been confirmed, in our investigation, as an appropriate
starting point for modeling U.S. foreign policy decision making. We would
argue further that the use of formal, computational modeling procedures
has allowed us to demonstrate and explore the utility of a cybernetic model
in a systematic, dynamic mode, not feasible without AI-based methodologies.
Further Considerations of the Use of Rule-Based

Models for the Study of Foreign Policy Making
While the assessments noted above may be taken as supporting the
general direction of our investigation and as indicating UNCLESAM's capacity
as a behavioral model, we conclude with some observations and reservations,
based on our experience with UNCLESAM, for the application of models
with simple rule-based architectures to the study of foreign policy decision
making. There is the matter of domain specificity, an often-cited limitation
of rule-based models. (See Mallery, 1987, and Mefford, this volume.) How
generalizable, then, do we view UNCLESAM? Our answer is qualified. First,
it is important to note that UNCLESAM's architecture (knowledge base,
PAL, rules, and metarules) are not idiosyncratic to the locale or the dates
of the Dominican Republic case. The model was designed to represent a
stereotypic cognitive structure and a stereotypic decision process, arguably
typical of U.S. images of, and policy making toward, countries in Central
America and the Caribbean. In this sense, it could be applied to the history
of U.S. relations with a series of countries over extended periods of time.
In fact, this is what would be required to make any strong assertions
concerning the breadth of the UNCLESAM's domain. Based on our limited

experience with the extended Dominican Republic case study, it is our view
that within the framework of the U.S. foreign policy of use of force toward
Central America and Caribbean states, UNCLESAM is likely to provide a
satisfactory representation of U.S. interaction with these regimes in the
post-World War 11 era. There is sufficient basis, in substantive treatments
of U.S. foreign policy, and in the relative success of UNCLESAM's application
to the Dominican Republic, to argue that the model's modified cybernetic
logic could apply to additional cases in this domain. 14
The designing and application of UNCLESAM, however, also demonstrated
in important ways the need for models with more sophisticated architectures
to further represent and explore the process of U.S. foreign policy making.
In other words, if one's ambitions are to get inside the "black box," simple,
rule-based models remain deficient in fundamental ways. Two examples,
drawn from UNCLESAM and the Dominican Republic case provide ample
support for this claim. First of all, there was in our model no capability to
capture and reflect the timing of U.S. foreign policy decisions or the time
delays associated with policy action. UNCLESAM was designed to react
immediately and chronologically to its input event stream. And while, in a
general sense, the model's response patterns concerning use of force
deployments were accurate, UNCLESAM could not cope with representation
of classes of events such as diplomatic maneuvers, economic policy steps,
and cooperation with other states. Policy decisions such as these appear
to have proceeded according to a different logic, i.e., not a simple, direct
action-response cybernetic logic. Such U.S. policy actions occurred without
apparent linkage to specific, recent events or changes in the Dominican
Republic-reaction times were varied and inconsistent; responses were
indirect rather than direct. Also, they appear to have involved multiple
decision agents within the United States (the Congress, the Pentagon, the
State Department, the president) and to have been contingent upon the
configuration of domestic considerations in the United States and upon U.S.
international relations with other states in and outside Central America.
The second caveat regarding UNCLESAM is illustrated, specifically, with
the matter of postures and posture shifts in UNCLESAM. But, in general
terms what is exemplified is the lack of cognitive content in the cybernetic
process of UNCLESAM's architecture. Recall that UNCLESAM operated
relatively successfully in a context-dependent rule application mode. Decisions
concerning the shifting of posture could not be handled by UNCLESAM;
the model had to be provided this information (context) as an input. The
context, i.e., determination of the particular sets of rules relevant in light
of U.S. attitudes toward the Dominican Republic, was dictated by the
specification of U.S. Posture at each particular time point. To cause UNCLESAM
to function over a broader range of policy outputs and at a higher level of
sophistication would require the model to be able to generate its own posture
shift decisions. Upon examination of both the historical record and foreign
policy analyses, we find, however, that this is a type of decision making
by U.S. policy-makers that could not adequately be encompassed by an
input-output model of a unitary actor decision-maker. This is, in other words,
a decision making about cognitive attitude changes-not representable by
a relatively simple "black-box" model such as UNCLESAM.
Complex architectures with alternative logics for decision making, e.g.,
script-based reasoning or case-based reasoning (see Mefford, this volume),
and more sophisticated representations of the numbers and interactions of
decision agents have been advanced within the artificial intelligence com-
munity and are beginning to be applied to the study of international relations.
Our subsequent work, for example, still focused upon U.S. behavior toward
Central America, is designed to explore the applicability of multiagent, script-
based architectures of decision making by foreign policy bureaucracies (Job,
Johnson, Selbin, 1987 and 1988). We view this work as continuing in
important ways the enterprise commenced with UNCLESAM. First it continues
a logic of inquiry that advances incrementally from simple to more complex
and more sophisticated formal models and accompanying theoretical ex-
plorations. Second, it recognizes the strengths and weaknesses of rule-based
logics and corresponding model structures, such as seen in UNCLESAM.
Notes
1. Elsewhere we have detailed both more preliminary (Job and Johnson, 1986)
and subsequent research results (Job, Johnson, and Selbin, 1987, 1988).
2. See Job, Johnson, and Selbin (1988) for a detailed elaboration of this argument.
3. Others besides Schoultz (e.g., Whitehead, 1983) have studied and described
U.S. decision making concerning Central America. While Schoultz's work reflects
the results of hundreds of interviews with policy participants, his characterization
of the policy process still emerges as among the more simply dimensioned. For our
purposes, Schoultz constituted a good benchmark for constructing our model, in
that refinements and complexities could be added as deficiencies in the model were
revealed through its failure to track adequately U.S. policy outputs.
4. Note that "instability for U.S. policy makers is defined in a quite straightforward
fashion as being equated to 'violence'" (Schoultz, 1987: 37, note 7).
5. The simplest cybernetic decision-maker operating in a fashion analogous to
the Watt governor or even in the manner of Ashby's homeostat, the cat who moves
in reaction to heat from a fireplace, neither has nor needs any memory (see
Steinbruner, 1976: 53).
6. UNCLESAM was written in INTERLISP and operated upon Xerox 1100 series
computers.
7. There are twenty-seven Action Categories in the PAL of UNCLESAM, including
behaviors such as investigate, assassinate, return, attack, use-force, threaten, and
request, as well as changes in levels of behaviors such as stop-action, change-action

level, and increment/decrement-action-level.
8. The PAL involves five Relations (diplomatic-relations, president, vice-president,
leader, and posture) and two Attributes (stability and ideology).
9. We do not underestimate the inherent difficulties of attempting to construct
an "accurate" record of U.S. perceptions of historical events, nor the amount of
intensive labor required to achieve a more satisfactory record, based on archival
sources. (See Majeski and Sylvan, this volume, for further discussion of these issues
as relevant to another project concerning U.S. foreign policy decision making.) For
exploration of the potential of our cybernetic-type model, use of newspaper report-
based input was satisfactory.
10. On the seven-point stability scale, a level of 7 would indicate conditions, as
perceived by the United States, of complete civil disorder1civil war in the Dominican
Republic. Note that in our view the "normal" level of stability conditions in the
Dominican Republic during this period was defined as 3. That is, in the absence of
reports of disturbance within the republic, stability levels in UNCLESAM were
constrained to go no lower than 3.
11. These were 7 = placement of U.S. military forces on DR territory, 6 =
military threats/actions by immediately proximate U.S. units, 5 = deployment of
significant military units immediately offshore, i.e., a show of force, 4 = deployment/
announcement of military units relevant for action in the DR. Lower values, 3-1,
referred to the readiness status of units on the U.S. mainland.
12. These actions could be viewed as a form of exogenous or "external" truth
maintenance. That they were necessary highlights a serious difficulty, both in practice
and in principle, for the modeling of real-world foreign policy decisions.
13. In a more finely grained assessment of UNCLESAM's replication capacity,
we examined each individual instance in which UNCLESAM produced a response
against the record of U.S. behavior. The degree of correspondence that one requires
to declare a match between U.S. behavior and UNCLESAM's output is important.
Also, one must distinguish between hits and misses, where the latter are separated
into false positives (i.e., UNCLESAM acts but the United States did not) and false
negatives (i.e., UNCLESAM actions do not correspond to those of the United States,
either by producing no action or by producing nonmatching actions). If an exact
match (in terms of both type, direction, and magnitude) is demanded, about two-
thirds of U.S. actions were matched by UNCLESAM's responses. But we would argue
that this imposes an artificially rigorous standard on UNCLESAM, given the simple
nature of the model and its limited response capabilities. If we utilize a modified
(and, we would argue, more appropriate) standard, where a match is made if
UNCLESAM produces a response when the United States acted, and if UNCLESAM's
response was of the same type and direction, then over 90 percent of U.S. actions
are reproduced by UNCLESAM. There were, however, several instances of false
positive results. These were, with the exception noted in the text, minor instances;
and with the occurrence of no further events in the Dominican Republic, UNCLESAM's
rules caused the model to revert back to a previous level. UNCLESAM could, therefore,
be assessed as being a bit "overly sensitive" to events in the Dominican Republic,
a not surprising result, given that the model was oriented around assessments of a
single indicator, instability, which it determined according to a quite broadly defined

set of categories of observations, e.g., riots, attacks on the opposition, etc., and
without a memory or capacity to take into account conditions or events outside the
Dominican Republic.
The analysis described above constituted a type of "sensitivity analysis," in that
we observed the model's response to a variety of input stimuli. We did not, however,
engage in extensive or systematic "sensitivity analysis," in the sense that this term
is used to refer to analysis of the relative sensitivity of the model across a range
of stimuli and/ or the relative impact of the various rules of the model in determining
model output. Nor did we engage in testing the model with "counterfactuals,"
following the example of Thorson and Sylvan (1982), among others.
14. On the other hand, it is unlikely that models like UNCLESAM would prove
useful starting points for exploring U.S. foreign policy toward the Soviet Union or
in regions in which great-power interests have consistently clashed. Modeling these
situations entails design of architectures involving numerous actors interacting in
strategic ways, with memories, learning capacities, etc.
Bibliography
Alker, H.R., Jr., and C. Christensen (1972) "From Causal Modeling to Artificial
Intelligence: The Evolution of a UN Peace-Making Simulation," in Experimentation
and Simulation in Political Science, J. Laponce and P. Smoker, eds. Toronto:
University of Toronto Press.
Etheredge, L. (1985) Can Governments Learn? American Foreign Policy and Central
American Revolutions. New York: Pergamon Press.
Job, B., and Johnson, D. (1986) "A Model of U.S. or Foreign Policy Decision Making:
The U.S. and Dominican Republic, 1959-1965," paper presented at the annual
meeting of the International Studies Association, Anaheim.
Job, B., Johnson D., and Selbin, E. (1987) "A Multi-Agent, Script-Based Model of
U.S. Foreign Policy Towards Central America," paper presented at the meeting
of the American Political Science Association.
___ (1988) "U.S. Foreign Policy Decision Making Concerning Central America:
Further Developments of a Multi-Agent, Script-Based Model," paper presented
at the meeting of the American Political Science Association.
Job, B., and Ostrom, C. (1985) "Opportunity and Choice: The U.S. and the Political
Use of Force, 1948-76," paper presented at the annual meeting of the International
Political Science Association, Paris.
LaFeber, W. (1983) Inevitable Revolutions. New York: Norton.
Lowenthal, A. (1972) The Dominican Intervention. Cambridge: Harvard University
Press.
Mallery, J. (1987) "Thinking About Foreign Policy: Finding an Appropriate Role for
Artifically Intelligent Computers," unpublished paper, MIT Department of Political
Science and Ai Laboratory.
Martin, J. B. (1966) Overtaken by Events. New York: Doubleday.
Ostrom, C., and Job, B. (1986) "The President and the Political Use of Force,"
American Political Science Review 80: 541-566.
Schoultz, L. (1987) National Security and United States Policy Toward Latin America.
Princeton: Princeton University Press.
Slater, J. (1978) "The Dominion Republic, 1961-66," in Force Without War: U.S.
Armed Forces as a Political Instrument, B.M. Blechman and S.S. Kaplan, eds.
Washington, D.C.: Brookings Institution, pp. 289-342.
Steinbruner, J. (1976) The Cybernetic Theory of Decision Making: Perception,
Cognition, and Artificial Intelligence. New York: Praeger.
Thorson, S.T., and Sylvan, D.A. (1982) "Counterfactuals and the Cuban Missile
Crisis," International Studies Quarterly (December 1982), 26(4): 539-571.
Whitehead, L. (1983). "Explaining Washington's Central American Policies," Journal
of Latin American Studies 15.
Wiarda, H. (1984) Rift and Revolution: The Central American Imbroglio. Washington,
D.C.: AEI Press.
10
Modeling Foreign Policy Decision Making
as Knowledge-Based Reasoning
Donald A. Sylvan, Ashok Goel,
and B. Chandrasekaran
Politics and Information Processing

In this chapter, we report on an experimental knowledge-based system
called JESSE, for Japanese Energy Supply Security Expert, 1 that models
some aspects of Japanese energy policy decision making (Goel and Chan-
drasekaran, 1987; Goel et al., 1987a; Sylvan et al., 1990). The system is
initiated by supplying information about an energy-related event, such as
the Iranian revolution of 1979. It recognizes the threat posed by the event
to Japanese energy supply security and delivers a set of plans appropriate
for the situation. In deciding on a set of plans, the system takes into account
the state of Japanese foreign relations that impose constraints on the choice
of policy options. Thus, JESSE performs the complex information-processing
task of constrained decision making, which involves the tasks of threat
recognition, constraint formulation, and reactive planning.
A fundamental premise is that political mechanisms are only one part
of a full account of the political decision-making process. Political actors
are also problem-solvers, who make decisions by exploring different choices,
who plan and execute various actions, who make use of their knowledge
This research has benefited from contributions by Davis Bobrow, Brian Ripley, Michael Weintraub,
and Jelfrey Morrison; we are grateful to them. In addition, this chapter has benefited from
discussions with Charles Hermann, Valerie Hudson, Dwain Melford, Richard Miller, Bradley
Richardson, Harvey Starr, and Stuart Thorson, as well as from three helpful meetings sponsored
by the Midwest Consortium for International Security Studies. We thank them. We acknowledge
the support of the Mershon Center of the Ohio State University, the U.S. Air Force Office of
Scientific Research (AFOSR-87-0090), and the U.S. Defense Advanced Research Projects Agency
(RADC-F30602-85-C-0010).
245
246 D. A. Sylvan, A. Goel, and B. Chandrasekaran
and experience in the pursuit of their interests and goals. These activities
can be viewed as information-processing phenomena.
As recognized by many political scientists, problem-solving agents are
only boundedly rational, being subject to a number of constraints on
computation and memory (Simon, 1969, 1985).
Decision-making behavior of political units, we argue, is the result of
interactions between these political and information-processing mechanisms.
Depending upon the norms of a political culture, the degree of consensus
among the actors on a given problem, and previous experience in a particular
problem domain, the decision-making process can be located along a
spectrum. At one extreme of this spectrum, decision making is dominated
by purely political mechanisms, while at the other extreme information-
processing mechanisms play an important role. The latter is exemplified by
highly organized problem-solving actors who accept politically dictated values
and goals and then carry out the decision-making process with little
interference from political mechanisms. Any specific political decision-making
instance represents a particular mix of political and information-processing
mechanisms, the latter always at the service of the former.
As one would expect, literature in political science contains a number of
proposals for studying political mechanisms-what they are and how they
work. However, there has been relatively little scholarship on the role of
information-processing mechanisms in political decision making. Of course,
the forms of these mechanisms are unlikely to be specific to politics.
However, the content of the information processing would be dictated by
the politics of the decision-making situation being modeled. As we will
illustrate in the substantive domain addressed by our model here, we argue
that there are many aspects of international relations in which information-
processing mechanisms play a central role.
Political Content and Information-Processing Languages

Over the last few decades, research in the field of artificial intelligence
has led to an information-processing language that offers new primitives
and constructs for capturing the problem-solving behavior of intelligent
agents. In this language, problem-solving agents achieve their goals by
representing problem spaces (i.e., the set of choices and possibilities), using
knowledge to explore these spaces, and choosing suitable alternatives (Newell,
1978). This language, as we will illustrate, provides a medium for expressing
theories of political decision making with greater precision and explicitness
than before.
Foreign Policy Decision Making as Knowledge-Based Reasoning 247
Political Content in the Information-Processing Model

What does the information-processing model JESSE have to do with the
political part of Japanese energy-related decision making? That is, where
is the "politics" in it?
The politics in the theory occurs at two levels. First, the knowledge
structures and the problem spaces-the classification hierarchies, the plans-
have politiCal content. Thus, any proposal in this language, including the
one realized in JESSE, provides a theory about how the problem-solving
mechanisms, internal politics, and international politics come together both
to create and to delimit the options.
The second level at which the theory has political content is the degree
to which the proposed model uses problem-solving mechanisms to the
exclusion of the political mechanisms. In fact, we chose the Japanese energy
policy as the domain precisely because this may be a decision-making
instance where the problem-solving components dominate the process. Japan
has so "rationalized" its decision making that politics at best provides
constraints for the space of alternatives (e.g., plans) and knowledge structures
(e.g., plan schemas) but does not enter in the form of political mechanisms.
Once goals and available options are politically set, the process is largely
one of problem solving rather than one of political pushes and pulls.
Alternative models proposing a collection of actors (e.g., interest groups)
could be modeled using JESSE-like constructs, but the interaction of the
actors is modeled by political mechanisms largely modeled in other languages
that have the expressiveness for the pushes and pulls of political accom-
modations. The attractiveness of the family of approaches embodied in this
chapter arises from new methodological possibilities for dealing with such
complex models.
Information Processing of Individuals and Groups

There are several goals to the JESSE project:
1. To demonstrate the usefulness of the information-processing perspective

by showing that it enables one to describe aspects of political decision
making that could not be considered in detail before, especially in
any formal manner. This contribution can be judged on the basis of
what aspects of the phenomenon can be made explicit using this
methodology, aspects that might remain hidden or vague without this
formalism.
248 D. A. Sylvan, A. Goe/, and B. Chandrasekaran
2. To demonstrate the expressiveness of a particular family of information-

processing languages called the generic task languages for capturing
many relevant aspects of problem solving and decision making by
political units. This family of languages is, however, not argued to be
complete in its present form. As more forms of information-processing
mechanisms for capturing problem-solving behavior become available,
the family can be extended with additional members.
3. To propose a specific theory of Japanese energy-related decision
making using the above language. The evidence for this theory will
be discussed, as will the issues of validation and scope conditions.
Performance and Validation

JESSE is not a performance system in two senses. In one sense, following
the competence/performance distinction that was originally made by Chom-
sky (1965), JESSE is a competence model, not a performance model. It
sets out the "virtual machine" that characterizes the options and the
knowledge that underlies the exploration of options but does not make a
commitment to a direct mapping between its parts and the decision-making
units in the real world of Japanese energy policy decision making. It is also
not a performance model in another sense, namely, it does not have sufficient
knowledge represented in it to test if its decisions correspond to the ones
that were in fact made by Japan in a wide enough variety of cases. Of
course, such a model would not be simply a validation of a theory but
could also be a potential source of information about future decisions by
the political unit being modeled. In other domains, investigators in AI have
constructed systems that play precisely these multiple roles.
The rest of this chapter is organized as follows: In the next section we
discuss our reasons for choosing the domain of Japanese energy supply
security and propose a political theory for Japanese energy and foreign
policy decision making in this domain. In the third section we present an
information-processing analysis of the decision-making task and propose a
functional architecture for the JESSE system. In the fourth section we
discuss the issues of knowledge, inference, control, and implementation of
JESSE, including the distributed artificial intelligence aspects of our model.
In another section we address the validation of the political information-
processing model embodied in JESSE. In the section after, we discuss the
scope conditions of the model. In another section we discuss our efforts in
the context of rule-based and case-based modeling efforts. The final section
offers conclusions.
The Domain of Japanese Decision Making
Choosing a Domain
Domains of political decision making that have the following properties
are appropriate candidates for the construction of models such as JESSE,
which require a significant expenditure of effort:
1. The domain is of sufficient importance to be worth the investment.

Japan is a country that is poor in natural resources such as energy
and is critically dependent on energy-exporting countries for its needs.
2. The knowledge requirements can be circumscribed, i.e., open-ended
world knowledge is not called for, or the problem can be decomposed
so that human problem-solvers can interactively solve certain sub-
tasks-in foreign policy problems (this is more realistic), when the
number of relevant international actors is relatively few in number.
We argue that these circumstances more accurately portray the Jap-
anese energy policy decision domain than they do other domains of
Japanese decision making.
3. The knowledge structures and strategies are sufficiently slowly changing
that they can be assumed fixed. Whether this property accurately
describes the Japanese energy policy domain is an empirical question.
The substantive content of our theory reflects our judgment that this
property is met.
Overall, Japanese energy policy making exhibits many elements of the

rational actor model. While the common U.S. view of "Japan, Inc." surely
overstates the point, Japanese government and business do not have a deep
institutional rivalry (see, for instance, Samuels, 1987). Especially in the
domain of energy supply security, Japanese policy decision making is not
subsumed by institutional or bureaucratic competition. Additionally, more
than is the case in many other countries, Japanese decision-makers have
similar political socialization patterns. The preponderance of Japanese civil
servants, for instance, have been educated at Tokyo University, with most
of the remainder having been educated at Kyoto University (see Richardson
and Flanagan, 1984; Kubota, 1969). The self-selection process for those
who want to be decision-makers leads to the study of law and economics
in quite a high percentage of cases. Generally similar foreign policy worldviews
is hardly a surprising result, and we characterize the reasoning of such an
elite as "group information processing."
Although these characteristics, in broad form, are not unique to Japan,

they raise interesting questions, which we address later, about the scope
conditions for generalizing from our model to decision making by political
and economic elite of other nation-states.
The availability of information about the actor being modeled and knowledge
about the domain being studied must also be considered. We garner our
information both from the wealth of literature on the subject of the Japanese
energy domain and from personal interviews with Japanese political and
economic elite.2 The academic, business, and Japanese governmental com-
munities, for example, have all produced informative literature.
A Political Analysis of Japanese Foreign

and Energy Policy Decision Making
Decision Making Under Constraint. Japan is a nation-state that has very
real resource constraints but succeeds more than do other nation-states by
approaching the constraints in a novel manner. This view serves as the
starting point for our analysis of Japanese decision making in the domain
of energy supply security.
Economically Centered Concepts of National Security. On a "substantive"
level, JESSE begins from the assumption that Japanese leaders are quite
concerned with national security but view the concept of national security
as a fundamentally economic one. Those who criticize Morgenthauian realism
(Morgenthau, 1966) as being too concerned with military factors to the
exclusion of economic influences would clearly receive a sympathetic hearing
among Japanese elite. One of the most significant observations stemming
from the initial interviews on which the "first cut" of this model was based
is that Japanese leaders tend to view "national security" as an economically
as opposed to a militarily centered concept. The JESSE model attempts
to draw out the implications of this distinction. As will be seen, the substantive
categories for classification in JESSE are such economic categories as
"energy cost" and "energy flow," rather than such military categories as
"nuclear weapons state" or "military ally." The political content of our
theory, then, is based on observing that Japanese decision-makers dealing
with foreign energy policy act more similarly toward nations that happen
to be energy suppliers, for instance, than they do toward nations that happen
to have similar military capabilities.
Anticipatory Policies. We posit that Japan has prepared in advance policy
options for anticipated threats to its energy supply security (Bobrow et al.,
1986; Sylvan et al., 1987). How does Japan select such an option in response
to an energy-related event? lt begins by choosing from a menu of these
anticipated policies. We have observed that certain domestic economic means
have been followed on a number of occasions. Those are found in the
"unilateral adjustments" portion of figure 10.1. Both political and economic

means, including concentrating on the United States, fall under the con-
ceptually distinct bilateral domain. Finally, a set of options that are mostly
economic in nature are stored in the multilateral portion of Figure 10.1.
Unilateral, bilateral, and multilateral options may be pursued simultaneously,
with a slightly more political emphasis-consistent with our understanding
and theorizing about Japanese energy decision making-in the bilateral
category.
State of Foreign Relations. Japanese energy policy decision making takes
place in the context of its foreign relations in general. At the time of the
Iranian revolution, for instance, Japan's relations with some Far East Asian
countries were strained. What is the role of Japanese foreign relations
generally in its energy policy decision making?
The specific aspects of Japanese foreign relations represented in JESSE
are Japanese relations with Far East Asian countries, Japanese-U.S. security
relations, openness and stability of the international economic order, U.S.
support for the international economic order, and access to foreign markets
for Japanese exports. Figure 10.2 illustrates these categories. Interviews
with Japanese leaders together with such literature as Samuels, 1987 and
(MITI, 1985) lead us to hypothesize that such a classification scheme is
consistent with Japanese leaders' conceptualization of that decision domain.
JESSE: Function and Architecture

The discussion thus far has been confined to purely political aspects of
Japanese energy policy making. We have not yet talked about the role of
information-processing mechanisms in the given decision-making instance.
The political aspects of the model provide only an incomplete and inexact
description of Japanese energy-related decision making. The model, as
explicated thus far, does not specify, for instance: (1) how the anticipatory
policies are stored, indexed, and retrieved; (2) how the constraints imposed
by the Japanese foreign relations situation are accommodated in the decision
making; or (3) how the economically centered concepts are represented,
organized, and used. Moreover, the concepts, as introduced thus far in our
discussion of the politics of the model, are specified by their label only
(e.g., "cost"); neither the content of these concepts nor the relation between
them is clear.
In information-processing terms, while the political aspects of our model
of Japanese energy-related decision making partially characterize the problem
space to be explored, they do not specify how this problem space is
represented and explored, what knowledge is used to explore this space,
and how this knowledge is represented and organized. Below, we present
an information-processing model that seeks to address some of these issues.
Figure 10.1: Planning Hierarchy
t:ll
[\)
[--- BuyEnergyShares I
SubsidizeDepletableResources
Unilateral I De~el~pRe,;;;;.,ableResources ~
Adjustments
ReduceEnergyDemand
StockpileEnergy
PurchaseEnergyEisewhere
EncourageUSAction
Anticipatory Bilateral
Policy Relations
AvoidAIIiances
I Induce Techn~~gic;,;d;;;nd-;,;;;8]
BolsterReliableExporters
SupportCo/lectiveAction
Multilateral
Management UseForeignShipping
[ SupportConsumerCartel I
[ FundlnternationaiR&D I
Figure 10.2: A Slice of the Classification Hierarchy 11
MajorProblemln
JapaneseAsian
Relations
Japanese CurrentProblemln
Asian JapaneseAsian
Relations Relations
Japanese MajorProblemln
us JapaneseUS
Security Security Relations
Relations
CurrentProblemln
JapaneseUS
Japanese International SecurityRelations
- Foreign J---~ Economic
Relations Order
Note, however, that the problem space has (by and large) already been
specified by the already explicated political aspects of the model. The
information-processing model embodied in JESSE accepts this politically
dictated problem space as a given.
Theory of Generic Tasks

Our information-processing analysis of Japanese energy-related decision
making is based on the theory of Generic Tasks (Chandrasekaran, 1983,
1988). This theory recognizes that complex problem-solving tasks, such as
political decision making, often are computationally very hard and expensive.
254 D. A. Sylvan, A. Gael, and B. Chandrasekaran
How can boundedly rational actors perform these complex tasks with limited
information, processing time, and memory space? The theory proposes that
complex tasks are often performed by decomposition into a small set of
generic tasks. A generic task is a "natural kind" of information-processing
task, corresponding to which is a strategy that provides a basic building
block of intelligence. Classification (Bylander and Mittal, 1986; Gomez and
Chandrasekaran, 1981), and Plan Selection (Brown and Chandrasekaran,
1985, 1986) are two examples of such generic tasks. A generic task is
specified by the information it takes as input and the information it gives
as output. The strategy corresponding to a generic task is characterized
by the primitive types of knowledge, inference, and control needed for a
computationally efficient transformation of the input to the output. Thus, if
knowledge is available in the required forms, then the strategy corresponding
to the generic task of classification ensures an efficient mapping of the
input (a description of a specific situation) to the output (concepts that
pertain to the specific situation) (Goel et al., 1987b).
Below, we first characterize the task of Japanese energy-related decision
making and then present a functional architecture for it, which uses generic
tasks as building blocks.
The Decision-Making Task

Characterizing a specific problem-solving task in Japanese energy and
foreign policy decision making in the domain of its energy supply security
is the starting point of our analysis. We argue that a goal of the Japanese
political and economic elite in the domain of Japanese energy supply security
is to maintain the state of low-cost supply of imported energy commensurate
to its energy needs. Often, this goal is threatened by the occurrence of
certain energy-related events in the world. An important task in maintaining
the goal state is deciding what actions to take in response to a given energy-
related event. The input information for this task is a description of the
Japanese energy situation following the energy-related event as well as a
description of the Japanese foreign relations situation at that time. The
output information is some subset of the set of anticipatory policy options
stored in memory.
The task of transforming the given input information into the desired
output information in Japanese energy-related decision making can be viewed
as a special form of routine planning. In routine planning (Chandrasekaran
et al., 1989), the problem-solving actor typically has access to partial plans
that are indexed over the specifications of anticipated goals. Given an
explicitly specified goal to be achieved, the actor uses the goal specifications
as indices to retrieve the appropriate partial plans and then deliberatively

synthesizes them into an overall plan that can accomplish the goal. In the
specific instance of Japanese energy-related decision making, however, the
appropriate stored plans are retrieved from memory in response to specific
situations, where a situation can be described in terms of certain states in
the environment of the problem-solving actor. We may call this form of
planning situation-reactive routine planning. In situation-reactive planning,
the goal of the problem-solving actor is not explicitly represented; instead,
it is implicit in the mechanism for retrieving the stored plans. Moreover,
there is no deliberative synthesis of the stored plans; instead, the actor can
potentially invoke each plan retrieved in response to a given situation.
A Functional Architecture
The next step is to analyze the functional architecture for computationally
efficient retrieval of stored plans from memory. In principle, it may be
possible to directly index the stored plans over specific states in the decision-
making actor's environment. However, since the number of relevant states
that may occur in the environment is potentially very large, a direct mapping
of the states onto the stored plans can be computationally very expensive
(Chandrasekaran and Goel, 1988). This mapping may be performed more
efficiently by decomposing it into two tasks. First, the relevant states in
the environment may be classified onto a small number of stored concepts,
where each concept represents an equivalence class of some subset of the
states. Next, these concepts may be used as indices for retrieving the plans.
Now, according to our political theory of Japanese energy-related decision
making, there are two kinds of states in the environment that are relevant
to the decision-making task: the states that describe the Japanese energy
supply situation and the states that refer to the Japanese foreign relations
situation. These two different kinds of states can be separately classified
onto concepts that are semantically relevant to them. Thus, the states
describing the Japanese energy supply situation can be classified onto
concepts that represent specific types of threats to the country's energy
security. Similarly, the states that describe the Japanese foreign relations
situation can be classified onto concepts that represent specific types of
problems in foreign relations. Next, the threats to Japanese energy supply
security and the problems in Japanese foreign relations can be further
classified into indexical categories. These indices may now be used for
selecting appropriate plans in storage.
This analysis leads to the functional architecture for the JESSE system
shown in Figure 10.3. The architecture contains four modules, three of
Figure 10.3: Functional Architecture of JESSE
A description of the A description of

Japanese Energy Japanese Foreign
Supply Situation Relations Situation
Classification Classification
Module I Module 11
Threats to lnternationa
Japanese Problems
Energy Supply 1 Facing Japan
Security
Classification
Module m
Complex Indices for

Plan Selection
Plan
Selection
Module
'
Energy-related Policies
which perform the generic task of classification, while the fourth performs
the generic task of plan selection.
Classification of the Japanese Energy Situation

Let us now consider the identity of the modules in the functional architecture
of Figure 10.3 and the rationale for their content more closely.
The first classification module takes as input a description of the Japanese

energy situation following an energy-related event and maps it onto hypotheses
about the threats posed to Japanese energy supply security. The organization
of these hypotheses in JESSE is shown in Figure 10.4. The label EnergyCost
in the figure refers to the hypothesis "an increase in the cost of energy."
Similarly, EnergyFlow refers to the hypothesis "a reduction in the flow of
energy." Note that this organizational scheme makes clear the relations
between the various concepts. Thus, FlowDueToChangelnExportCapability
("a reduction in the ftow of energy due to a decline in the energy export
capability of some country") is a subhypothesis of EnergyFlow, and a
superhypothesis of CostDueToChangelnExportCapability ("an increase in
the cost of energy secondary to a reduction in the ftow of energy due to
a decline in the export capability of some country").
We note that CostDueToChangelnExportCapability is in the EnergyFlow
branch of the tree in Figure 10.4 instead of the EnergyCost branch. This
is because an increase in the cost of energy to Japan can occur in two
ways. First, the cost of energy can increase, for instance, if some energy-
exporting country decides to increase the price per unit of energy it exports
to Japan. Second, the cost of energy can increase, for instance, if an
energy-exporting cartel decides to reduce the amount of energy it exports.
Of course, once it has been judged that there is indeed a cost threat, then
the response to it is likely to be the same irrespective of whether the treat
is of the first or the second kind. For this reason, we collapse the two
kinds of cost concepts in the first classification module into a single indexical
category in the third classification module of JESSE.
Selection of Stored Plans

The second classification module of Figure 10.3 takes as input a description
of the Japanese foreign relations situation and maps it onto hypotheses
about various types of problems in Japanese foreign relations. The organization
of these hypotheses is the same as in the previous section. The conceptual
subdivisions of the slice of the second classification module in Figure 10.2
find their genesis and rationale in the type of terms often used by Japanese
leaders in expressing options and classifying problems they face.
The third classification module of Figure 10.3 takes as input the hypotheses
established in the first classification module as well those established in the
second classification module, and maps them onto indexical categories. The
organization of these indexical categories in JESSE is shown in Figure 10.5.
The label BothEnergyAndOtherlssuelmportant refers to the index of "there
is a serious threat to Japanese energy security and there is also a serious
problem in Japanese foreign relations." The thesis is that Japanese response
to an energy-related event depends on the relative importance of the threat
(3{
eo
Figure 10.4: A Slice of the Classification Hierarchy I
FlowOueTo Immediate Cos roue To

SupplyRouteCutoff Change In
ExportCapability
HowOueTo CostOueToChangeln
SupplierBoycorr ExportCapability
MajorCostDue To
FlowOueTo Change In
Change In lmmediateCflangeln ExportCapability
ExportCapability ExportCapability
FlowOueTo MajorChangeln
Change In ExportCapability
ExportPolicy
Immediate Cost
Major Cost
Figure 10.5: A Slice of the Classification Hierarchy Ill
Energy
Issue
Present
Neither Both
Energy Energy
Nor And
Other Other
Issue Issue
Important Important
Other Energy
Issue Issue
Dominant Dominant
Energy Energy
Cost Flow
Issue Issue
Dominant Dominant
Major Immediate Major Immediate

Energy Energy Energy Energy
Cost Cost Flow Flow
Issue Issue Issue Issue
Dominant Dominant Dominant Dominant
that the event poses to energy supply security and on the problems in
Japan's foreign relations. That is, Japanese response to a threat to its energy
supply security varies depending on whether Japan faces an even more
pressing foreign relations problem at that time. In the extreme case, if the
energy threat is minor and not immediate while the foreign relations problem
is major and current, then Japan might not devote much immediate attention
to the threat. This explains why OtherlssueDominant has no subconcepts
in the classification hierarchy.
The fourth module of Figure 10.3 takes as input the indexical categories
established in the third classification module above and gives as output an
appropriate subset of the set of anticipatory policy options. We have already

described these anticipatory policies in the previous section (see Figure
10.1 ). Here we would like to address two aspects of our political theory
incorporated in Figure 10.1: why a distinction between unilateral, bilateral,
and multilateral, and why these plans as opposed to others.
We see Japanese foreign policy decision-makers as having prepared in
advance for many contingencies, that is, many plans are "stored." The
interviews on which we based the initial cut of JESSE led us to believe
that many Japanese decision-makers viewed options according to whether
and with whom they had to deal or coordinate. The "unilateral, bilateral,
multilateral" distinction follows from that feature of the initial interview
responses. Plans from each of these three categories can and are at times
pursued simultaneously in JESSE.
Why we chose the particular set of general plans illustrated in Figure
10.1 is a question with a similar answer. Those initial interviews and our
understanding of Japanese reasoning in the energy supply arena consistently
point to many of the specific contingent options listed on the right side of
Figure 10.1. For example, time and again Japanese elite have referred to
"stockpiling energy," "inducing technological dependence," and "using foreign
shipping." We do not claim that we have captured all stored plans, only
that there is evidence that the plans summarized on the right side of Figure
10.1 seem to be developed and "on the shelf" for future use by Japanese
foreign policy decision-makers.
Foreign Policy Decision .Making as Knowledge-Based Reasoning
JESSE: Knowledge Representation and Inference Mechanisms

In this section, we discuss the representation of the concepts, and the
inference mechanisms and control of processing in JESSE.
Classification of Japanese Energy Supply

Each of the hypotheses in the first classification module in JESSE is
really a schema that is represented as a frame (Minsky, 1975). The frame
corresponding to the hypothesis of FlowDueToChangelnExportCapability
(i.e., "reduced energy flow to Japan due to a change in the export capability
of some energy-exporting country") specifies its superhypothesis as
EnergyFiow, and its subhypotheses as .MajorFlowDueToChangelnExport-
Capability, lmmediateChangelnExportCapability, and CostDueToChange-
lnExportCapability.
Each such frame also contains knowledge in the form of questions about
the energy situation that JESSE asks of the user, for the purpose of deciding
whether the hypothesis of FlowDueToChangelnExportCapability is established

as being plausible for a given energy situation. Thus, when the hypothesis
of FlowDueToChangelnExportCapability is considered, JESSE asks the user
about the decline in the energy production and transportation capabilities
of the energy-exporting country in question.
The knowledge for combining the answers to these questions into decisions
about the plausibility of the hypothesis is represented in the form of production
rules (Newell, 1973). Thus, the first rule in such a frame combines the
answers to the questions regarding decline in the production and transportation
capabilities of an energy-exporting country into a plausibility value for the
hypothesis that there is a threat of reduced energy flow to Japan. Note
that it is possible for the user to answer JESSE's questions with "unknown."
JESSE accommodates the incompleteness of input information by choosing
a lower plausibility value for the hypothesis. If the plausibility value in the
hypothesis is high enough ("likely," or higher), then the hypothesis is
established; if the plausibility value is in intermediate range, then the
hypothesis is uncertain; otherwise the hypothesis is rejected.
Finally, the frame contains control knowledge that specifies what to do
if the hypothesis of FlowDueToChangelnExportCapability is established, and
what to do if it is not established. If the hypothesis is established, then it
is refined; the hypothesis sends invocation messages to its subhypotheses,
and transfers the control to them. Now, the above process repeats itself at
the next level in the classification hierarchy. If, however, the hypothesis is
not established, then it is not refined any further. Instead, the control of
processing is returned to its superhypothesis, in this example to EnergyFlow.
This strategy for classification has been called establish-refine (Gomez and
Chandrasekaran, 1981).
Preparation of Complex Indices

The knowledge representation and inference mechanisms in the second
and third classification modules of JESSE are the same as in the first
classification module; each concept is represented as a frame that contains
rules for combining evidence and determining plausibility values. The control
of processing is also the same: establish-refine.
The plausibility value for each concept in JESSE is stored in a data
object (a list) that is placed on a shared memory. Thus, the plausibility
value of a concept in one module is accessible to other concepts in the
same module, as well as to concepts in other modules. The default plausibility
value for each concept is set very low (e.g., "very unlikely"). As the
plausibility value of various concepts in the different modules of JESSE are
determined, these values are updated. The schema corresponding to each
indexical category in the third classification module contains knowledge
about what concepts in the first and second classification modules are
relevant to it and how to combine their plausibility values into a confidence
factor for that index.
All three of the classification modules in JESSE have been implemented
in the CSRL language on a Xerox 1108 workstation. CSRL (for Conceptual
Structures Representation Language) (Bylander and Mittal, 1986) is a high-
level knowledge representation language that embodies the strategy of
establish-refine. CSRL itself is implemented in lnterlisp-D/LOOPS, where
lnterlisp-D is a version of the LISP language and LOOPS is an environment
that supports object-oriented programming.
Selection of Stored Plans

We turn now to the issues of knowledge and control in the fourth module
of the JESSE system. The control of processing that JESSE uses for the
task of plan selection can be called instantiate-expand. First, each plan (i.e.,
each anticipatory policy option), starting from the top node in the hierarchy
of Figure 10.1, uses the indices established in the third classification module
to check whether it is suitable for a given energy supply and foreign relations
situation. If so, the plan is instantiated (or sponsored); if not, it is rejected.
Second, the sponsored plan is expanded by considering lower-level plans
in the hierarchy, which repeat the process.
Associated with each plan sponsor is a frame containing knowledge in
the form of production rules that determine, based on the confidence factors
of the indices, whether or not the plan is suited to a given situation. Thus,
when the plan EnergyShares is considered for sponsorship, the confidence
factors of the relevant indices are fetched from the shared data structures,
and the suitability of the plan is determined. If the plan is found to be
suitable, then it is sponsored, automatically instantiated, and, if possible,
expanded.
The module for plan selection has been implemented in the DSPL language,
on a Xerox 1108 workstation, as were the classification modules. DSPL
(Design Specialists Planning Language) is a knowledge representation language
that supports the strategy of plan selection and expansion (Brown and
Chandrasekaran, 1986). Like CSRL, DSPL is implemented in Interlisp-D/
LOOPS.
Distributed Information Processing

Information processing in our model is distributed at two different levels
of abstraction:
1. Information processing is distributed functionally over the multiple

problem-solving and planning organizations that cooperate in the
decision making. There are three organizations that perform the task
of hierarchical classification, and one that performs the plan selection
and refinement task.
2. Information processing is distributed conceptually within each functional
organization. In three of the organizations, a community of problem-
solving agents cooperatively performs hierarchical classification. Sim-
ilarly, in the fourth organization a community of planning agents
cooperatively performs plan selection and refinement.
Validation of the Model

Validating complex information-processing models such as JESSE is
fraught with methodological problems. We have sympathy for the argument
that there is no "objective reality" against which to validate a model. Rather,
each actor and analyst understands the world independently. Our standard
of validation, then, is a comparison of JESSE with understandings of
Japanese decision making that have been expressed publicly.
One problem in undertaking this type of validation is that the language
of classification and plan selection is almost too expressive from the viewpoint
of validation. There are many ways in which these languages could be used
to fit the data points, and it would be difficult to argue, at this state of the
implementation, that JESSE has been comprehensively compared to any
alternative representation. However, were we to have found that any of the
four validation steps to be described below led us to judgments that JESSE
was operating implausibly, we would have pointed out the disparities to the
reader and altered the model. This did not, in fact, happen.
The problem of validation is not a logical problem but one of size. One
could ask of JESSE what JESSE would recommend as action in various
situations. However, in order to really validate JESSE, the scope of JESSE
would have to be increased considerably, i.e., a much larger infusion of
knowledge would have to be given to JESSE. This is an expensive undertaking,
possible in principle but likely in practice only after the plausibility of models
such as JESSE is demonstrated, and that, too, only for problems whose
importance or significance can justify the cost of the knowledge-engineering
effort.
With these caveats in mind, we have undertaken the following steps:
1. We have tested our model against widely shared representations of

outcomes of some recent situations. Two examples are the Iranian
revolution, and the removal of Sheik Yamani from the post of the oil
and petroleum minister of Saudi Arabia. Our results show that the
performance of JESSE is reasonable, meaning that the outcomes of
JESSE are generally in line with what is publicly represented as having

transpired. Nothing resembling "contradictory" circumstances took
place. Although buoyed by such results, we should point out that
building a performance system is not our major objective; we are
more interested in understanding the process by which national elite
arrive at policy decisions. We do acknowledge, though, that the
understanding we achieve is related to the performance of the system.
2. We have tested our model on hypothetical situations also. An example
is the hypothetical situation in which Indonesia and Malaysia are at
war, and Malaysia has threatened to close the Strait of Malacca to
all international shipping. As with others who have undertaken coun-
terfactual analysis (e.g., Thorson and Sylvan, 1982), checking process
validity has been our focus here.
3. We have demonstrated the system to a few domain experts. This,
too, has been an attempt to check the process validity of our model.
Their judgment so far has been that the energy policy decision-making
process followed by JESSE is plausible.
4. We have conducted a literature survey to determine if there is some
evidence that literature represents Japan as following the energy
decision-making process modeled by JESSE. Japanese language doc-
uments (e.g., MITt, 1985) are part of this survey. Since the model
itself is based upon interviews with Japanese political and economic
elite, we are not checking the model against information from which
we built it. Our literature "tests" suggest that Japan does indeed
classify energy-related events onto the types of threats that they pose
to its energy supply security and does select and refine stored plans.
Also, the temporal ordering of information sought by the Japanese
government and business interests suggested by JESSE fits with such
documents in the few cases in which information that specific is
available. For example, queries about the feasibility of collective action
seem to have been made only after the collection of flow and cost
information. This temporal ordering is consistent with the process
posited in JESSE.
We feel that for empirical validation of a model such as ours, the tests that
we have just described are more appropriate than statistical tests. One
reason for this conviction is that our model allows for such a broad base
of multiple outcomes. In other words, Japan, in our model, can undertake
no actions in response to an external event, or it could undertake a dozen
or more actions, simultaneously, some of which would seem contradictory.
Therefore, statistical tests such as those offered by Bueno de Mesquita
(1981) are inappropriate.
Additionally, each of the four tests outlined above examine both outcome
and process validity. Our position is that we offer this model into the
academic debate concerning how decisions, including Japanese decisions,
are made. We claim neither that it is the only true model nor that it is the
best. We do, however, claim that it illuminates aspects of decision making
that other efforts have not done.
Scope Conditions of the Model
Generalizability
When considering the issue of generalizability, it is important to note
that what we see as the core of the model is the way in which information
is processed, and not the substance of the plans in the planning section
of the model. In other words, while our vision of progress in science is not
in full agreement with Lakatos (1970), we see the "hard core" of our theory
as the notion and the mechanisms of information processing, not as particular
plans or actions that the model predicts. Thus, we believe that the idea of
modeling decision making by focusing on information processing can gen-
eralize to all decision domains (with varying degrees of utility as noted
earlier in this chapter), and modeling group information processing can
generalize to at least the domains that have these four charac;teristics:
homogeneous elite political socialization, sacrosanct domain, limits on in-
stitutional rivalry, and evidence of prior planning.
In such cases, some of the fundamental conftictual elements of politics
have been resolved either prior or exogenously to the onset of the decision
domains in question. Goals are often quite clear in this subset of decision
environments, oftentimes because they include maintenance of what are
perceived to be essential functions of the polity. As a result, the process
of group decision making unfolds almost as though an individual were
processing information.
Extensions to the Model

As we see it, JESSE presently stands on it own as a plausible information-
processing model of Japanese energy decision making. In the future, we
hope to further improve the model. The two directions for further refinement
and improvement that we anticipate are as follows:
1. As we have mentioned earlier, one of the tasks that JESSE performs

is reactive planning. Reactive planning is event-driven rather than goal-
directed; there are no explicitly represented goals in JESSE. We believe
that along with reactive planning Japanese energy policy decision
making also involves maintenance planning. In maintenance planning

the goals of maintaining certain functional states in a stationary state,
for instance, keeping the Strait of Hormuz open to international shipping,
are explicitly represented. Maintenance planning involves the task of
goal identification, which uses a hierarchy of goals. We have developed
a preliminary design for maintenance planning.
2. We are working toward augmenting JESSE with a data base that
allows for knowledge-directed data abstraction and inference. The
current version of JESSE lacks this capability. Thus, JESSE may
acquire from the user knowledge regarding an energy shortage in the
world energy markets, and later may need to know if there is an
energy glut but cannot infer it from prior knowledge. An intelligent
data base would alleviate this problem.
Relation to Rule-Based and Case-Based Approaches
Rule-Based Decision Models

At first glance it might appear that the theory of generic tasks for
modeling political decision making is similar to that of rule-based systems
(see Mefford; Job and Johnson, this volume), since a portion of JESSE's
knowledge is represented as production rules as we described earlier. In
fact, there are several important differences between the two approaches:
• First, our analysis of Japanese energy-related decision making is at the

level of the generic information-processing tasks it performs. Since the
language of production rules is at a lower level of abstraction altogether,
it does not offer any constructs for capturing task-level distinctions,
as JESSE does.
• Second, the knowledge in JESSE is organized around concepts that
are represented as frames. The concepts themselves are organized into
hierarchies. This leads to a more "molecular" form of knowledge
representation and information processing. Rule-based systems, in con-
trast, have a more "atomic" representation of knowledge.
• Third, the semantics of the concepts is different in the various modules
of JESSE: The concepts are hypotheses in the first and second
classification modules, indexical categories in the third module, and
plans in the fourth module. Again, the language of production rules
offers no primitives to capture these semantic differences at the con-
ceptual level.
• Fourth, the generic task framework explicitly specifies the abstract
control of processing for each generic task. For instance, the abstract
control structure is establish-refine for the classification task and in-
stantiate-expand for the task of plan selection. In contrast, in rule-

based systems, the higher-level control is implicitly embodied in a
control formalism at a lower level of abstraction, which is uniform
across tasks.
• Finally, the generic task methodology enables more efficient processing,
easier knowledge acquisition, and more perspicuous explanation of
knowledge and processing; Chandrasekaran (1987) provides a more
detailed discussion of these issues.
Of course, it is possible to implement JESSE as a rule-based system. The

point, however, is that the theory of generic tasks offers a higher-level
vocabulary, which allows the political theorist to capture certain functional
and conceptual distinctions that might remain opaque in the language of
production rules.
Case-Based Decision Models

Case-based reasoning (Schank and Abelson, 1977; Schank, 1982; Mefford,
this volume) offers a different approach to modeling decision making. In
this approach, decision-making instances are stored in memory as cases.
When the decision-making actor is faced with a new decision-making situation,
the cases that best match the description of the situation are retrieved from
memory. The actor's response to the new situation is derived by analogy
with the previously encountered situations.
Case-based reasoning offers a computationally attractive alternative to
approaches that emphasize reasoning from "first principles" for every new
decision-making situation. We are in general agreement with this approach;
however, we would like to argue that cognitive agents store their experiential
knowledge not only in the form of cases in memory but also as more
abstract compiled knowledge. They do so because it is typically compu-
tationally more efficient to use knowledge in compiled form whenever it is
available and applicable.
The compiled structures in Jesse are the concepts, the indices, and the
plans stored in memory. When JESSE encounters a new decision-making
situation, it classifies the description of the situation, establishes appropriate
indices, and instantiates the stored plans that are suitable to the situation.
We do not see this approach to modeling decision making as inconsistent
or incompatible with that of case-based reasoning. While JESSE is not a
a case-based reasoner, it would be very interesting and useful to investigate
how knowledge gets compiled from individual cases into the kind of structures
present in JESSE. Clearly, it is not feasible to store all information about
a previous situation, the actor's response to it, and the consequences of
his actions, because that may require an extremely large amount of memory
space. Instead, it would appear that intelligent agents typically process and
interpret raw information about a case, prepare complex multiple indices
for it, and then store only those aspects of the case that they have some
reason to believe might be especially useful in the future.
Whether or not experiential knowledge can or should be operationalized
and compiled into higher-level structures depends on a number of factors:
• Nature of the Task: Abstraction of operational knowledge from expe-

riential knowledge often requires a relatively large sample of cases over
which generalization can occur. If the nature of the task is such that
the agent has encountered few cases of it, then the agent may not
have a large enough sample to abstract operational knowledge. Thus
operational knowledge typically is available in compiled form only for
tasks that are relatively routine. We note that the task that JESSE
performs satisfies this condition. Since the world energy situation has
been volatile for many years, Japan has had to deal with a large number
of energy-related problem situations and thus has acquired a large
repertoire of cases. Thus it has had a large enough sample of individual
cases from which operational knowledge can be abstracted and encoded
into compiled structures.
• Nature of the Domain: Abstraction of operational knowledge from
experiential knowledge is typically computationally very expensive. Thus,
even if a large enough sample of individual cases were available, an
agent may not generalize over them if there were no special reason to
do so. For instance, if the domain of the cases is of little interest or
importance to the agent, then he may not devote the computational
resources required for abstracting operational knowledge and encoding
it into compiled structures. The domain of Japanese energy supply
security, of course, is of critical importance to Japan, since continued
Japanese economic prosperity depends on a steady flow of energy at
reasonable cost. Thus Japan has had ample reason to devote much
attention to previously encountered energy-related situations, the re-
sponses to them, and their consequences, and to learn from them.
• Nature of the Agent: Even if the task is relatively routine and the
domain is important, an agent might not generalize over individual
cases if it lacks the required computational resources. Often the
computational resources required for operationalization of experiential
knowledge are so enormous that they are beyond the capacity of a
single individual. In such case, groups of individuals may collectively
abstract the appropriate operational knowledge for encoding into com-
piled structures. Often this may take the form of concurrent, distributed
processing involving a number of agents. Each agent may view a
relatively small sample of cases from its own local perspective, and
arrive at partial generalizations. The group may collectively complete

the generalization process by some form of negotiation between the
individual agents. We note that it is quite possible that the individual
agents in the group may be unable to resolve their differences by
negotiation (or any other process), in which case there may be no
commonly agreed-upon generalizations. In the case of Japan, it appears
that there exists a consensus among its elite at least in regard to policy
decision making in the domain of energy supply security. Thus the
Japanese elite as a group of intelligent agents appears to satisfy the
conditions under which generalization over individual cases can occur.
We have so far argued that (1) it is computationally advantageous to encode

operational knowledge into high-level compiled structures, and (2) the
conditions under which experiential knowledge can be operationalized and
encoded into compiled structures are apparently satisfied in the case of
Japanese energy policy decision making. In JESSE the compiled structures,
stored in the system's memory, are the plans and the complex multiple
indices used to select the plans. When it encounters a new problem situation,
it classifies the description of the situation and matches it with the complex
multiple indices. The plans whose indices best match the description of
the situation are invoked. Thus JESSE identifies some of the plans and
indices used in Japanese energy policy decision making and offers a functional
architecture for retrieving the plans from memory. In this sense, JESSE
can be viewed as a compiled case-based reasoner.
Conclusion
JESSE is a significant research endeavor because it has attempted to
represent an understanding of decision making without modeling the behavior
of specific institutions or of specific individual decision-makers. Despite that
(and we would argue that it is in fact because of that), the information-
processing approach or metaphor incorporated in JESSE has allowed us to
capture Japanese behavior in quite a plausible manner.
Though we have presented evidence for the plausibility and validity of
JESSE, the amount of work that would be required to give definitive evidence
on behalf of information-processing models such as JESSE is substantial.
JESSE would have to be tested on an even larger body of cases than we
have subjected it to thus far, which in turn would require a substantially
enlarged knowledge base. However, we feel we have convincingly argued
that information-processing methodology is a potentially powerful meth-
odology, which should be considered carefully in attempting to understand
political decision making.
Notes
1. While the "Expert" fits well as part of our acronym, the rationale for our
research is not building a usable expert system for policy making, though that may
be a useful by-product. Instead, we seek to construct, test, and refine an information-
processing theory of foreign policy decision making.
2. Interviews on which the first cut of our model was based were conducted by
Davis B. Bobrow, who collaborated with us on the early stages of this project.
Bibliography
[Aiker, 1975] Hayward Alker. "Polimetrics: Its Descriptive Foundations," in Fred
Greenstein and Nelson Polsby: editors, Handbook of Political Science, Volume 7,
Reading, Massachusetts: Addison-Wesley, 1975, pp. 140-210.
[AIIison, 1971] Graham T. Allison. Essence of Decision, Boston: Little, Brown, 1971.
[Bobrow et al., 1986] Davis Bobrow, Donald Sylvan, and Brian Ripley. "Japanese
Supply Security: A Computational Model." Presented at the annual meeting of
the International Studies Association, Anaheim, March 25-29, 1986.
[Brown and Chandrasekaran, 1985] David C. Brown and B. Chandrasekaran. "Plan
Selection in Design Problem Solving," in Proceedings of the Artificial Intelligence
and Simulation of Behavior Conference, University of Warwick, United Kingdom,
February 1985.
[Brown and Chandrasekaran, 1986] David C. Brown and B. Chandrasekaran. "Knowl-
edge and Control for a Mechanical Design Expert System," IEEE Computer,
19(7):92-100, July 1986.
[Bueno de Mesquita, 1981] Bruce Bueno de Mesquita. The War Trap, New Haven:
Yale University Press, 1981.
[Bylander and Mittal, 1986] Tom Bylander and Sanjay Mittal. "CSRL: A Language
for Classificatory Problem Solving and Uncertainty Handling," A/ Magazine, 7(3):66-
77, August 1986.
[Chandrasekaran, 1983] B. Chandrasekaran. "Towards a Taxonomy of Problem-solving
Types," A/ Magazine, 4(1):9-17, 1983.
(Chandrasekaran, 1986] B. Chandrasekaran. "Generic Tasks in Knowledge-based
Reasoning: High-level Building Blocks for Expert System Design," IEEE Expert,
1(3):23-30, Fall 1986.
[Chandrasekaran, 1987] B. Chandrasekaran. "Towards A Functional Architecture for
Intelligence Based on Generic Information Processing Tasks," in Proceedings of
the Tenth International Joint Conference on Artificial Intelligence, Milan, Italy,
August 1987, pp. 1183-1192.
[Chandrasekaran, 1988] B. Chandrasekaran. "Generic Tasks as Building Blocks for
Knowledge-based Systems: The Diagnosis and Routine Design as Examples,"
Knowledge Engineering Review, 3(3):183-219, 1988.
[Chandrasekaran and Goel, 1988] B. Chandrasekaran and Ashok Goel. "From Numbers
to Symbols to Knowledge Structures: Artificial Intelligence Perspectives on the
Classification Task," IEEE Transactions on Systems, Man, and Cybernetics,
18(3):415-424, May-June 1988.
[Chandrasekaran et al., 1988) B. Chandrasekaran, John Josephson, Anne Keuneke,

and David Herman. "Building Routine Planning Systems and Explaining Their
Behavior," in International Journal of Man-Machine Studies, 30: 377-398, 1989.
[Chapman, 1984) Daniel Chapman. "U.S.S.R.-Japan Energy Cooperation in Siberia:
Implications for U.S.-Japanese Relations," in C. K. Ebinger and R. A. .Morse:
editors, U.S.-Japanese Energy Relations: Cooperation and Competition, Boulder,
Colorado: Westview Press, 1984.
[Chomsky, 1965) Noam Chomsky. Aspects of the Theory of Syntax, Cambridge,
.Massachusetts: .MIT Press, 1965.
[Ebinger, 1984) Charles Ebinger. "U.S.-Japanese Nuclear Energy Relations: Prospects
for Cooperation/Conflict," in C. K. Ebinger and R. A. .Morse: editors, U.S.-Japanese
Energy Relations: Cooperation and Competition, Boulder, Colorado: Westview Press,
1984.
[Gaucas and Brown, 1987) Dale Gaucas and Alien Brown. "A Role for Assumption-
based and Nonmonotonic Justifications in Automating Strategic Threat Analysis,"
in Proceedings of DARPA Knowledge-based Systems Workshop, St. Louis, .Missouri,
April 21-23, 1987, pp. 158-165.
[Goel and Chandrasekaran, 1987) Ashok Goel and B. Chandrasekaran. "JESSE: A
Conceptual .Model of Political Decision .Making," in Proceedings of the Meeting
of the Midwest Artificial Intelligence and Cognitive Science Society, Chicago,
Illinois, April 23-24, pp. 133-135.
[Goel et al., 1987a) Ashok Goel, B. Chandrasekaran, and Donald Sylvan. "JESSE:
An Information Processing .Model of Policy Decision .Making," in Proceedings of
Third Annual Expert Systems in Government Conference, IEEE Computer Society,
Washington, D.C., October 19-23, 1987, pp. 178-187.
[Goel et al., 1987b) Ashok Goel, N. Soundararajan, and B. Chandrasekaran. "Complexity
in Classificatory Reasoning," in Proceedings of Sixth National Conference on
Artiliciallntelligence, AAA/, Seattle, Washington, July 14-18, 1987, pp. 421-425.
[Gomez and Chandrasekaran, 1981) Fernando Gomez and B. Chandrasekaran. "Knowl-
edge Organization and Distribution for .Medical Diagnosis," IEEE Transactions on
Systems, Man, and Cybernetics, 11(1):34-42, 1981.
[Goodman, 1982) Herbert Goodman. "Japan and the World Energy Problem," in D.
I. Okimoto: editor, Japan's Economy: Coping with Change in the International
Environment, Boulder, Colorado: Westview Press, 1982.
[Hiragawa, 1976) Kazushige Hiragawa. "Japan's Tilting Neutrality," in J. C. Hurewitz:
editor, Oil, the Arab-Israeli Dispute, and the Industrial World: Horizons of Crisis,
Boulder, Colorado: Westview Press, 1976.
[Krasner, 1972) Stephen D. Krasner. "Are Bureaucracies Important? (or Allison
Wonderland)," Foreign Policy, Number 7, pp. 159-179, Summer 1972.
[Kubota, 1969) Akira Kubota. Higher Civil Servants in Post-War Japan: Their Social
Origins, Educational Backgrounds, and Career Patterns, Princeton: Princeton
University Press, 1969.
[Lakatos, 1970)1mre Lakatos. "Falsification and the .Methodology of Scientific Research
Programmes," in lmre Lakatos and Alan .Musgrave: editors, Criticism and the
Growth of Knowledge, Cambridge, England: Cambridge University Press, 1970,
pp. 91-195.
[Lenat et al., 1983) Douglas Lenat, Albert Clarkson, and Garo Kiremidjian. "An Expert
System for Indications and Warning Analysis," in Proceedings of the Eighth
International Joint Conference on Artificial Intelligence, Karlsruhe, West Germany,
August 1983, pp. 259-262.
[Mefford, 1987) Dwain Mefford. "The Policy Process and the Evolution of Argument.
Casting Theory in the Form of a Series of Computer Programs." Presented at
the annual meeting of the International Studies Association, Washington, D.C.,
April 15-18, 1987.
[Minsky, 1975) Marvin Minsky. "A Framework for Representing Knowledge," in Patrick
H. Winston: editor, The Psychology of Computer Vision, New York: McGraw Hill,
1975, pp. 211-277.
[MITI, 1985) Ministry of International Trade and Industry. "Energy Policy White
Paper," Tokyo: MITI, 1985.
[Morgenthau, 1966) Hans J. Morgenthau. Politics Among Nations, Fourth Edition,
New York: Knopf, 1966.
[Newell, 1973) Alien Newell. "Production Systems: Models of Control Structures," in
W. Chase: editor, Visual Information Processing, New York: Academic Press, 1973,
pp. 463-526.
[Newell, 1978) Alien Newell. "Reasoning, Problem Solving and Decision Processes:
The Problem Space as a Fundamental Category," in Raymond Nickerson: editor,
Attention and Performance VIII, Hillsdale, New Jersey: Lawrence Erlbaum, 1978,
pp. 693-718.
[Read and Rosson, 198X) Steven Read and Mary Beth Rosson. "Rewriting History:
The Biasing Effects of Attitudes in Memory," Social Cognition, 1 (3) 240-255,
1982.
[Richardson and Flanagan, 1984) Bradley M. Richardson and Scott C. Flanagan.
Politics in Japan, Boston: Little, Brown, 1984.
[Samuels, 1987) Richard J. Samuels. The Business of the Japanese State: Energy
Markets in Comparative and Historical Perspective, lthaca: Cornell University
Press, 1987.
[Schank, 1982) Roger Schank. Dynamic Memory: A Theory of Learning in Computers
and People, New York: Cambridge University Press, 1982.
[Schank and Abelson, 1977) Roger Schank and Robert Abelson. Scripts, Plans, Goals
and Understanding, Hillsdale, New Jersey: Lawrence Erlbaum, 1977.
[Simon, 1969) Herbert Simon. The Sciences of the Artificial, Cambridge, Massachusetts:
MIT Press, 1969.
(Simon, 1985) Herbert Simon. "Human Nature in Politics: The Dialogue of Psychology
with Political Science," American Political Science Review, 79(2):293-304, June
1985.
[Sylvan, 1987) Donald Sylvan. "Supplementing Global Models with Computational
Models: An Assessment and an Energy Example," Behavioral Science, 32(3):212-
231, July 1987.
[Sylvan et al., 1987) Donald Sylvan, Davis Bobrow, and Brian Ripley. "Economic
Security as Managed Interdependence: A Computational Model of Japanese Energy
Policy." Presented at the annual meeting of the International Society of Political
Psychology, San Francisco, July 4-7, 1987.
[Sylvan et al., 1990) Donald Sylvan, Ashok Goel, and B. Chandrasekaran. "Analyzing
Political Decision Making from an Information Processing Perspective: JESSE,"
American Journal of Political Science, 34(1):74-123, February 1990.
[Thorson, 1984) Stuart J. Thorson. "Intentional lnferencing in Foreign Policy: An AI
Approach," in Donald A. Sylvan and Steve Chan: editors, Foreign Policy Decision
Making: Perception, Cognition, and Artillciallntelligence, New York: Praeger, 1984.
[Thorson and Sylvan, 1982) Stuart J. Thorson and Donald A. Sylvan. "Counterfactuals
and the Cuban Missile Crisis," International Studies Quarterly. 26(4):539-571,
December 1982.
11
Decision Making and Development:
A ''Glass Box'' Approach to Representation
Margee M. Ensign and Warren R. Phillips
The litany of problems experienced by many developing countries in the

past decade has become all too familiar. The debt burden of the LDCs,
which reached $1.3 trillion in 1988, overshadows attempts to reignite growth
and undermines political stability, particularly in the fragile democracies of
Latin America. In Africa, the debt-service ratio in 1988-1990 for the poorest
countries is projected to reach 30 percent, a figure that represents an
intolerable and unsustainable burden. Real commodity prices for many of
the LDCs' primary exports have declined since the mid-1970s. The future
growth of markets for many of these exports is questionable as substitutes
like fiber optics are developed for copper, and as the Western world turns
away from the tropical oils. Environmental degradation in many of the LDCs
is threatening short-term growth prospects and the future of sustainable
development. Finally, after some dramatic gains in health, education, and
life expectancy in the 1960s, problems of hunger and poverty are once
again widespread.
The field of development studies itself is experiencing an upheaval as
traditional assumptions about growth and development are reexamined. 1 This
chapter highlights an important area being overlooked both by economists
and political scientists-political decision making. It describes an artificial
intelligence-based simulation that is being developed, 2 and concludes with
a discussion about language and theory.
Political Development and Decision Making

Many developing countries, particularly the big debtors, have followed
the structural adjustment programs put forth by the IMF, and yet growth
is still flat in these same countries. Worsening terms of trade, increasing
protectionism in many of the industrialized countries, and slower demand
274
A "Glass Box" Approach to Representation 275
accounts for only part of the stagnation. Why are these programs successful
in some developing countries and not in others? Why, despite slower growth
and tightening markets, have some countries been able to keep development
plans on track?
One answer to these questions, and an area that is missing in the debates
among economists and political scientists, is an analysis of political decision
making. Particularly in the developing countries, politicians intervene in
markets for a variety of reasons having to do with their image of how
development should proceed and the need to reward supportive or threatening
interest groups in order to ensure political stability, if not survival.
Decision-makers in the LDCs are constantly faced with critical economic
and political choices. They must decide how to trade off politics against
economics. They must resolve the pressure to engage in long-range planning
versus short-term reactions to problems. They must respond to pressures
from the international arena and balance these pressures against domestic
demands. Finally, they must decide how to focus their attention on different
segments of society.
Decision-makers do not make decisions about these major directions
daily, but they do have a mental perspective, a cognitive preference, in each
of these areas. Because when people are satisfied with their material well-
being they are less likely to antagonize the government, decision-makers
can build or restructure coalitions and increase or decrease centralization
with only broadly imposed constraints. That is, they can follow their cognitive
preferences without concern for the stability of their government.
But if development is insufficient, there is likely to be conflict arising
from numerous sources. Decision-makers are well aware that a rapid rate
of economic growth may secure their political future. Thus, there is con-
siderable motivation to accelerate growth as rapidly as possible. Unfortunately,
LDCs will only rarely have substantial enough growth rates to supply this
needed security. As noted, particularly since the late 1970s, growth has
been flat in many LDCs. As a result, decision-makers are beseeched by
various groups wanting to at least maintain existing benefits from the
government. These benefits are often inconsistent with the goal of increasing
growth. In a cycle that constantly repeats itself, decision-makers must yield
to groups that could provide political support, and because of this, they
are prevented from obtaining the desired growth rate. Political survival,
therefore, means that decision-makers must cultivate political support groups
(such as the military) that they can satisfy. It becomes important to these
groups to keep the government in power, even if this may mean political
suppression.
However, the picture is more complex than even this scenario. Other
political processes penalize governments for coercive activity, centralization,
and, in the new democracies, for weakening democratic institutions. The
276 Margee M. Ensign and Warren R. Phi/lips
strength of these penalties is determined by country-specific cultural and

social mores. The overall political climate is partly determined by the relative
importance of constraints and incentives in these areas.
Our research is directed at understanding and representing the interaction
between decision making and policy in developing countries. We do not
dispute the importance of economic modeling, which most development
studies emphasize, but the importance of political decision making in guiding
development has gone unrecognized relative to the attention given structural
characteristics.
Two basic strategies guide our efforts:
1. We seek to relate political and economic questions.

2. We attempt to represent salient aspects of the information handling,
information interpretation processes of the key groups within specific
developing countries.
Most previous attempts to model development have been preoccupied

with prediction based on structural characteristics and have missed the
dynamics of adaptation. The recognition, however, of the complexity, in-
terdependence, and long lead times involved in attempting to deal with the
problems of human and material resources has led to a number of attempts
to model complex social systems-the so-called global models. Prominent
examples include the Meadows's Limits to Growth, Forrester's World Dy-
namics, and Mesarovic and Pestel's Mankind at the Turning Point. The
causal relationships underlying these models, however, are based on the
interactions of aggregate variables: The human actors who shape a country's
development-their goals and plans-are missing from these models. These
types of models are unable to deal with human capability to cope with and
adapt to the future. The difference between the global models and the work
described in this chapter is clear: Instead of predicting the values of variables,
we are interested in understanding and representing the processes that lead
to specific decisions. In order to understand the difference between our work
and previous writers in development, it is useful to briefly review the literature
on political development and economic development, specifically focusing
on the difference between structure and performance.
Political structures refer to the persistent framework of relationships or
rules through which governmental action takes place. Thus the relationships
between the center and the periphery, the formal organizations of the
government, parties and special interest groups, the homogeneity or het-
erogeneity of the political culture, and the characteristics of the bureaucracy
are all structural components of political development. These components
usually change slowly, but they can also shift dramatically if some crucial
thresholds are reached. This view of reality is that which is represented in

the global models mentioned above.
Holt and Richardson (1969) define the performance of political systems
in terms of their problem-solving capabilities. The more developed a political
system, the greater its capacity to achieve a wide range of goals in a variety
of political and economic environments. Indeed, the process of political
development builds that capacity. That is, as nations develop, their experience
in successfully coping with problems increases their capability to deal
successfully with subsequent problems. One should find that a nation's level
of development is very closely tied to its problem-solving capability. This
argument is supported by numerous other scholars, such as Almond (1971),
Almond, Flanagan, and Mundt (1973), Pye (1965), and Binder (1970). They
argue that shifts in political structure, environmental challenges, and the
ways in which elite groups respond to these changes will have significant
effects on the subsequent capability of the political system to respond to
other challenges.
Structure and performance are also portions of the economic system.
Economic structure refers to the organization of the factors of production
and the distribution of consumption. For example, the interdependencies
between sectors and the mechanisms by which goods are distributed are
parts of the economic structure. Economic performance is simply the
productive output or the consumption level of the economy.
When examining the development process, most studies hold constant
the structures, both political and economic, and attempt to identify the
effects of one performance measure on another performance measure. They
may also concentrate on the interrelationship between various structural
variables such as the impact of economic productivity on political instability.
Each of these studies has made a timely addition to our knowledge about
some processes or has helped in the understanding of national performance
in one context. They are, however, inappropriate in explaining what appear
to be very similar development circumstances. We believe that the reason
these models do not work well as prescriptions for action stems from a
fundamental misrepresentation of development processes; the simplicity of
the causal models employed by analysts of development to explain societal
problems.
Moreover, the notion of the the rational actor has dominated much of
the debate in the social sciences. The rational actor approach, however,
cannot account for the complex social environment of development decision
making. This environment is characterized by an ever-changing array of
external and internal events. When we look at developing countries as they
struggle with the problems of reorganization, reform, choice, and, for some,
survival, one point stands out: Decision opportunities are fundamentally
ambiguous stimuli.
A major implication of established views of development was to see

political development in light of stages through which all systems must
pass. For years, scholars in development took on the task of looking at
what was termed the "underdeveloped" or "backward" countries and pres-
cribing what institutions and cultural traits of the West each lacked. The
state was viewed as essentially a distributional mechanism. That is, it had
or ought to have the capacity to distribute the goods of society in a manner
consistent with the need to industrialize. These authors saw in the history
of development the need for certain concomitant or prior conditions to
successful development. In other words, they "rationalized" the process of
development. If like Apter (1968) we view the world as system-bound, then
these approaches represent an absolute view of the nature of the problem,
and Huntington's argument for better management of the bureaucracy in
developing countries seems reasonable. Indeed this "rational" perspective
on managing development has become a consistent theme of the World
Bank recently.
The final strand of established thought that might be considered for
understanding and representing decision-makers in developing countries is
the literature on ·political decision making usually applied to bureaucracies
in the industrialized countries. The notions of incrementalism, satisficing,
and muddling through are also, however, based on the notion of the rational
chooser, which assumes that decision-makers have complete information
on all their alternatives along with the knowledge of the consequences of
their choices and the ability to rank-order preferences. None of the literature
discussed thus far, whether development studies or research on bureaucracies,
has seriously questioned the rationality assumption.
Recent work by cognitive scientists has cast much doubt on the traditional
conception of rationality, however. In fact, one of the most significant ideas
to emerge from the field of cognitive science is the importance of perception
and context in human thought, behavior, and decision making. Like other
authors in this volume, we have drawn on this information-processing
approach to decision making and used some of the major findings as the
basis for our representation. Specifically, two assumptions underlie the
development of our models: that prior beliefs and experience within an area
are important factors in determining future actions, and that decision-makers
develop analogies to deal with similar policy decisions.
We have chosen developing countries where a strong leader dominates
the national scene (Egypt, Nigeria, Kenya, Tanzania) and Jamaica, where a
leader with a specific ideological orientation (Seaga) has been replaced by
one with an opposing ideology and set of policies (Manley). We are attempting
to model the images or cognitive maps of these leaders in areas where
they dominate the policy domain. This first stage of modeling involves
building the domain as it is seen from the leader's viewpoint by drawing
on memoirs, speeches, and key policy documents. This will give us a model
that corresponds with the viewpoint of the leader-our case.
The next stage of the modeling involves building in alternative perspectives
to model what could happen if the leader is confronted with a new problem
in the same policy domain. We argue that a leaders' prior beliefs and past
experience in dealing with a similar situation shape the perception of the
problem, and that based on these prior beliefs and past experience, an
analogy is developed to solve the problem. Finally, we are also interested
in examining the impact of alternative development strategies. By using the
same "initial conditions" facing the developing country leader we pose the
same problem-but allow the user to choose different policy options and
examine the impact of these choices.
To represent the knowledge in our simulations we have drawn on the
ideas of case-based reasoning and hypotheticals. 3 A real "case" in our model
is one in which a decision has been made by the leaders of the countries
being modeled: Egypt, Nigeria, Kenya, Tanzania, and Jamaica. The underlying
assumption is that when faced with a new situation, the leader-the model-
interprets the problem and develops solutions based on precedents. lt then
evaluates the solution using an interpreter that analyzes why a new situation
should or should not be classified in a certain manner. Using a past decision,
or a previous line of reasoning, the system evaluates the appropriateness
of a new solution. The system goes though several tasks when confronted
with a new situation:
• retrieving relevant cases from memory

• selecting the most appropriate case
• constructing an interpretation for the new situation or case
• evaluating the output from step number three
• storing the new case
This final step, storing the new case and indexing it in memory is potentially
very promising, given the concerns raised in this paper. This final step
involves, in a very rudimentary way, system learning. As Alterman argued,
"this step, the knowledge acquisition step, is the one that results in learning.
If a case was adapted in a novel way, if it was solved using some method
other than case-based reasoning, or if its results came from a combination
of results from several previous cases, when it is recalled during later
reasoning, the steps required to solve it won't have to be repeated. "4
The hypotheticals in the model are "possible worlds" generated from the
specific cases. As discussed by Rissland and Ashley (1986) the hypotheticals
are used as a heuristic technique. For example, in the simulation of Nigeria's
debt renegotiations, the specific case represents the Babangida's negotiations
with the IMF and the London Club (which deals with medium- and long-
term commercial debt) and Paris Club (which deals with rescheduling of
official debt). The model of the case allows the user to learn about the
specific actors and issues involved in this negotiation.
For example, Nigeria ran into severe balance of payment problems in
the early 1980s. Like like many of the oil-exporting developing countries,
Nigeria did not initially alter domestic policies to adjust to this change in
its external accounts. Government spending remained high despite declining
foreign trade receipts. By 1983 the economic situation was extremely serious,
discussions with the IMF had reached an impasse, and the regime in power,
the Shagari regime, was considered corrupt and inefficient. The Buhari
regime took power at this point, committed to contracting the domestic
economy rather than signing a new agreement with the IMF. Oil prices,
however, continued to drop, and existing domestic measures were insufficient
to deal with the problem. The Buhari regime was weakened because of its
increasing repression and rigidity. Major General Babangida took power in
a coup from Major General Buhari in August of 1985.
Our model begins at this point and describes how through early 1988,
Babangida dealt with the debt problem by reaching agreements with the
London Club, the Paris Club, the World Bank, and the IMF and how this
regime stayed in power by maintaining contact with the army and labor.
Then five alternative scenarios-our hypotheticals-along with the likely
consequences of each, are generated. The hypotheticals are alternative
scenarios with actions and likely outcomes along a continuum from a
continuation of the status quo through crisis management and an extreme
case. In the Nigerian case three hypothetical responses to the IMF agreement ,
reached in 1988 are: (1) the regime undertakes the reforms; (2) the regime
undertakes no significant reforms and attempts to handle the situation
through economic crisis management; and (3) a new regime comes to power.
Likely consequences from each of these scenarios are generated.
Once this stage of our modeling is complete, the final stage of model
building will involve including in our models the other relevant actors (both
national and international) and institutions (such as the IMF, political parties)
for each of the policy domains. What is novel about the system is not its
method of representation (for case-based reasoning programs have been
developed in the legal and medical fields, for example), but its accessibility.
In developing this aspect of the system, approaches and developments from
the AI and education field have been adapted.
Artificial Intelligence and Education

Much of the work of those developing At-based educational programs
has recently been devoted to developing intelligent computer-assisted in-
struction (ICAI) packages. With the introduction of At-based expertise into
computer-assisted instruction (CAI) systems in the late 1970s, the learning

environments of these educational programs became more oriented toward
discovery and problem solving.
These newer systems not only encourage interaction and discovery but
they are able to model the students' responses and provide important
feedback. Some of the more interesting examples of these types of programs
include SOPHIE, a program that teaches electronics; SCHOLAR, a geography
program; WEST, which teaches elementary mathematics; and BUGGY, which
can pinpoint students' misconceptions about arithmetic. 5
These systems have three major components: an expert module that
represents the expertise of the subject matter; a student module that attempts
to determine the user's moves and strategies so that "mistakes" can be
pointed out and corrected in the third module, the tutoring module. This
final module integrates knowledge about the subject area, specific pedagogical
goals, and natural language. 6
While these programs are major improvements over earlier CAI systems,
both their subject matter and their use is limited. Most of the ICAI packages
have been developed for the sciences and mathematics, domains where
knowledge is more complete and less controversial than in the social sciences.
These systems do, however, contain many important elements that should
be a part of ICAI systems developed for the social sciences. The computer
simulations of leaders in developing countries described above have drawn
on the significant developments in ICAI and have gone several steps further
than existing systems.
The simulations will contain two of the three modules. The system will
contain both an expert module containing "expert" knowledge about the
issue areas, and a student module to monitor the students' choices. This
module will not be used as the foundation for a tutoring module as is done
in most ICAI systems. Instead this module will be used to allow the user
to peer into the system and trace the causal connections that have generated
the simulated scenarios and that lead to changes in the scenarios. This
transparency is what Goldstein and Papert (1975) have called "the glass
box approach" toward representation. According to these authors:
The ICAI tutors, no matter how intelligent, are essentially black boxes from
the students' perspective. The possibility exists that the concepts developed
to design these programs might themselves be worthy objects of study by
the students. We propose calling this a glass box approach. 7
Because the simulations use this approach, at all points during the
simulation, the user will be able to "see into" the simulation and examine
the causal links, or the context of decision making that has led to the
particular scenario the user is examining and the effect that different policies
have on that scenario. The user can retrace the entire sequence of events
that have changed the scenario or can examine the impact of individual
choices on the scenario.
The system will not have the third module, the tutoring module, because
our knowledge of the issues simulated is controversial and far from complete.
But what we can do with the information available on these topics is to
make our assumptions, our theory, of decision making explicit. By so doing,
perhaps users will develop alternate theories and explanations-one of the
best outcomes that could occur as a result of using this system.
The simulation will encourage problem solving by allowing the user to
choose from a number of different perspectives and strategies in examining
and solving the problem or scenario presented. The user will also learn that
there is a wide range of possible responses to many different types of policy
interventions and that these responses are generated by context: That is,
perceptions of a problem and solutions generated to solve problems are
shaped by the domestic and international context of decision making.
Language and Theory

As this book demonstrates, the fields of artificial intelligence and cognitive
science offer a number of important ideas and methodologies that are relevant
to researchers in the field of international relations. Advances in AI and
cognitive science have opened up new ways for social scientists to understand
and represent knowledge. As a result, there is a rich storehouse of concepts
and tools that social scientists, particularly those working in international
relations, can draw upon to increase understanding and representation of
human behavior and interaction. Important areas of cross-fertilization have
occurred in the fields of decision making, problem solving, learning, and
representation. As a result of this interchange between political scientists
and scholars in these two fields, we have learned a great deal about how
humans process information, solve problems, and make decisions.
While real advancements in political science are occurring as a result of
this interchange, it is obvious to most of us working in this new area of
endeavor that few outside of our narrow circle are aware of or interested
in these efforts. There are at least three explanations for this state of affairs:
• the language gap between AI/IR and traditional approaches

• the inaccessibility of our models
• the impact of the dominant theory in the field-realism-on research
in international relations
The Language Gap

There are some parallels between the introduction of mathematical methods
in political science in the 1950s and 1960s and the current situation of
adapting artificial intelligence concepts and techniques. Like the period in
the 1960s when quantitative methods were first being introduced into political
science, only a small group of "experts" had sufficient mathematical back-
ground to understand and evaluate these methodologies. Certainly, under-
standing the various Al-based methodologies can be daunting. Moreover,
while the AI/IR community is at present a very small group, a surprising
diversity of concepts and methodologies from AI have been adapted to work
in international relations.
In discussing this earlier period in political science, Sartori thought that
the field had lost a "common mainstream of discourse," 8 that because of
the need by scholars to be perceived as having developed "new" or "original"
ideas (what Sartori called the "frenzy of novitism"), we ended up simply
playing musical chairs with words.
Like in this earlier period, the concepts and language of AI can be
baffling: terms like "object orientation," "scripts," "frames," "rules," "fuzzy
logic," and "intelligent information processing," draw, at best, blank stares.
Some lack of interest in our work clearly can be attributed to a "language
gap" between those using the language and concepts of AI and those
drawing on more traditional approaches in international relations.
In 1975 Sartori thought that the field was undergoing not a scientific
revolution in the Kuhnian sense, but a verbal revolution. He likened the shift
to a game of cards:
The relation of the linguistic instrument to scientific knowledge can be

visualized-with due account of some differences-as how the cards relate
to a card game. The game can be played only because the cards and the
rules of their combination are given-indeed static. Not too dissimilarly, only
a disciplined use of terms and procedures of "composition" and (decomposition)
permits the scientist to play his game. lt strikes me instead that we are
investing more and more of our energies simply in reshuffling and altering
the cards. 9
Sartori has a point: We must be careful in deciding which concepts and

which methodologies are most useful and appropriate from a political
scientist's point of view, and our language needs to clarify rather than
mystify. But in order to judge whether we are doing more than simply
playing musical chairs with words, it is critical that our theories of decision
making, cognition, and learning and our methods of representation are
accessible to the wider community of scholars in international relations.
Unlike the period of the 1960s, when the tools being introduced did not
question the dominant theoretical approach to the field, the approach of
scholars in AI/IR challenges some of the fundamental assumptions of realism
or its most recent version, structural realism. But it is impossible to evaluate
these theories unless our models, our programs, are accessible. This is
because our theories, our assumptions about behavior, are embedded in our
computer models. It is ironic that while we claim to be opening up the
"black box" of decision making, our own models remain in a black box.
We have had to take on faith that our models do the things we say they
do.
Although opening up our models and making our language clearer would
go a long way to making our work more accessible, by far the biggest
impediment to our research being examined or evaluated by scholars in
international relations is the hold that realism and now structural realism
has over the field.
Realism and Behavior

Structural realism ignores most of the evidence that has emerged about
human behavior-the focus of much of the work in AI/IR. Realists view
the state as a unitary, rational actor that acts and interacts in an anarchic
environment. The anarchic characteristic of the international system con-
ditions state behavior. From Thucydides through Morgenthau and Waltz,
realists (and structural realists) have taken an approach that views states
as rational actors-with consistent, ordered preferences-who maximize
their utility based on these preferences. Morgenthau and other realists state
that this description of states as rational actors is only an assumption that
allows their theories to be constructed: that these assumptions are neither
true nor false, only useful or not useful in constructing theory. Morgenthau
in Politics Among Nations states it is only an assumption that makes the
theory of international relations possible and that it "can be tested against
the actual facts. " 10
Waltz, in The Theory of International Politics, also makes this point: "We
can freely admit that states are in fact not unitary, purposive actors. States
pursue many goals, which are often vaguely formulated and inconsistent.
They fluctuate with the changing currents of domestic politics, are prey to
the vagaries of a shifting cast of political leaders, and are influenced by
the outcomes of bureaucratic struggles. But all of this has always been
known and it tells us nothing about the merits of balance-of-power theory." 11
It is possible, as we have seen from the field of economics, to predict
fairly accurately without necessarily understanding or explaining the phe-
nomena under study. However, the purpose of theory is to understand,
explain, and predict a chunk of the world. The rationality assumption,
whether for individuals or nation-states, has been seriously undermined by

scholars in cognitive science (see Purkitt, this volume).
These researchers and others have pointed out the importance of the
way in which a problem or decision is "framed," or the manner in which
the context of decision shapes that decision. Instead of rationally based
behavior, it appears that decision making is rule based, that is humans use
heuristics or rules of thumb in interpreting situations and making choices.
Much of the evidence from AI and cognitive science supports the view of
decision making described by Majeski:
Policymakers do not maximize. They do not evaluate all alternatives or calculate

expected utilities. They do not use rigid unchanging organizational standard
operating procedures. In order to deal with complex problems, policymakers
learn from experience, remember, and package their knowledge in the form
of heuristics and decision rules. These rules are verbal. Policymakers take
verbal knowledge (memory, current information, goals, ideological beliefs) and
use verbal rules to generate behavior (decisions, recommendations and the
like). 12
If much of the evidence on decision making supports the view that

behavior is rule following, that our images of the world shape our thoughts
and decisions, that the manner in which a problem or idea is "framed"
shapes our choices, why has this information had such a little impact on
mainstream political scientists? The research program, or indeed paradigm,
of realism has had a stranglehold on research and teaching in international
politics. This theory has held sway, except during periods where hostility
and conflict in the international system was in (at least temporarily) a decline.
The work of those focusing on interdependence, for example, began to be
noticed during the 1970s during a lull in superpower hostilities. It is entirely
possible, now that Cold War tensions seem to be abating, that scholars will
once again reexamine the critical assumptions underlying realism and ask
how useful they continue to be in interpreting the world and solving the
puzzles of international politics. 13 We may find that as we move into the
twenty-first century many of the assumptions underlying the realist tradition
not only are false but are no longer even useful in developing theory-in
understanding, explaining, and predicting the behavior of international actors,
whether they be states, organizations, or individuals.
But before we know how much the subfield of AlfiR can contribute to
the building of theory in international affairs, our models, our theories, must
be accessible for evaluation and critique. In "Judgment, Policy and the
Harmony Machine," John Fox states that AI could be a liberating force that
could change our understanding of and competence in the processes of
social judgment. Programs that are accessible, that give us the line of
reasoning, according to Fox also would be accountable systems. Until our

programs, our theoretical approaches, are transparent, however, it appears
as if we are only reshuffling the cards.
Conclusion
The goals of the models described in this chapter are twofold: to improve
our understanding of the decision-making processes of leaders in several
developing countries by building computer models that represent our theories
of these processes, and to use these simulations in educational settings.
One of the major difficulties for those of us working within the AI/IR
community is that few outsiders are able to examine our work. By building
a system that can be examined by other political scientists and used in
the classroom, we are attempting to open up our work to outside review
and use some of the promising techniques from AI to represent specific
cases of interest to students of international affairs. We are not claiming
our models can predict the future. We see our models as a very preliminary
step in increasing our understanding of decision making in developing
countries and our ability to represent these processes.
Notes
1. See Bruton, 1983; the entire issue of World Development, Vol. 11, No. 10,
1983; and Ensign, 1988, for a review of the major debates.
2. This project has been funded by the Ford Foundation. The eo-principal
investigators are Margee Ensign and Dr. Alan Aronson. Dr. Warren Phillips conducted
the background research for the Egyptian simulation, and Dr. Cheryl Christiensen
did similar work for Nigeria. George Shambaugh at Columbia University has been
a research assistant on the project. lt is anticipated that the simulations will be
completed in late 1991. The countries and issues being simulated are:
Egypt-structural reform
Nigeria-international debt
Kenya and Tanzania-agricultural development
Jamaica-debt and structural adjustment.
3. See Rissland and Ashley, 1986; Kolodner, 1987; and Hammond, 1986.
4. Richard Alterman, "Case Based Reasoning," Proceedings of the Case Based
Reasoning Workshop. Pensacola Beach, Florida, 1989, p. 5. See also Kolodner, 1985.
5. See Barr and Feigenbaum, Vol. 11, 1982, for the pedagogical orientation underlying
many of these programs.
6. Barr and Feigenbaum, Vol. 11, p. 233. Also see Goldstein, 1976, for an example
of an interesting tutoring strategy called· "coaching."
7. lra Goldstein and Seymour Papert, "Artificial Intelligence, Language and the
Study of Knowledge." MIT AI Memo, 337. July 1975, p. 65.
8. Giovanni Sartori, The Tower of Babe/: On the Definition and Analysis of

Concepts in the Social Sciences, Occasional Paper No. 6, International Studies
Association, University of Pittsburgh, 1975, pp. 8-9.
9. Ibid., p. 10.
10. Hans Morgenthau, Politics Among Nations, cited in Robert Keohane, ed.,
Neorealism and Its Critics. New York: Columbia University Press, 1986, p. 12.
11. Kenneth Waltz, "Anarchic Orders and Balances of Power," in Robert Keohane,
ed., Neorealism and Its Critics. New York: Columbia University Press,1986, p. 118.
12. Stephen Majeski, "A Recommendation Model of War Initiation: The Plausibility
and Generalizability of General Cultural Rules," in Stephen Cimbala, Artificial Intel-
ligence and National Security. Lexington, Massachusetts: Lexington Books, 1987,
pp. 62-63.
13. See Toulmin, 1963.
Bibliography
Almond, Gabriel. 1971 "Approaches to development causation," in Binder, Crises
and Sequences in Political Development. Princeton: Princeton University Press.
Almond, G.A., S.C. Flanagan and R.J. Mundt. 1973 Crisis, Choice and Change.
Boston: Little, Brown and Company.
Alterman, Richard. 1989 "Case Based Reasoning." Proceedings of the Case Based
Reasoning Workshop. Pensacola Beach, Florida.
Apter, David. 1968 The Politics of Modernization. Chicago: University of Chicago
Press.
Barr, Avron and Edward Feigenbaum. 1982 The Handbook of Artificial Intelligence.
Vol. 1-3. Los Altos, California: William Kaufmann.
Binder, Leonard. 1971 Crises and Sequences in Political Development. Princeton:
Princeton University Press.
Bruton, Henry. 1983 "The search for a development economics." World Development
Vol. 13, Nos. 10-11: 1099-1124.
Deutsch, Kart. 1953 Nationalism and Social Communication: An Inquiry into the
Foundations of Nationality. New York: John WHey and Sons.
Easton, David. 1965 The Political System: An Inquiry into the State of Political
Science. New York: Alfred Knopf.
Ensign, Margee. 1988 "Conditionality and foreign assistance: trends and impacts."
Journal of International Affairs (Fall 1988): 147-163.
Fox, John. 1985 "Judgment, Policy and the Harmony Machine." Proceedings of the
Ninth International Joint Conference on Artificial Intelligence. Los Angeles, Cal-
ifornia, pp. 1284-88.
Goldstein, lra. 1976 "The Computer as Coach: An Athletic Paradigm for Intellectual
Education." MIT AI Lab, Memo Number 389.
Goldstein, lra and Seymour Papert. 1975 "Artificial Intelligence, Language and the
Study of Knowledge." MIT AI Lab, Memo Number 337.
Haas, Ernest. 1975 "On Systems and International Regimes." World Politics Vol. 27
No. 2.
Hammond, Kristian. 1986 "Case-Based Planning: An Integrated Theory of Planning,

Learning and Memory." Ph.D. thesis, Yale University, Department of Computer
Science.
Holt, Robert and John Richardson. 1969 "Applications of automation theory to the
explanation of political development." Paper presented at the American Political
Science Association.
Kolodner, Janet. 1987 "Extending problem solver capabilities through case-based
inference." Presented at the Fourth International Workshop on Machine Learning.
lrvine, California.
Kolodner, J.L., Simpson, R.L. and Sycara, K. A. 1985 "Process Model of Case Based
Reasoning in Problem Solving." Proceedings of the International Joint Conference
on Artificial Intelligence. Los Angeles, California, pp. 284-290.
Moulaert, Frank. 1979-1980 "On the nature and scope of complex conflicts." Papers
of the Peace Science Society Vol. 29, pp. 49-66.
Parsons, Talcott and Edward Shills, eds. 1951 Toward a General Theory of Activity.
Cambridge: Harvard University Press.
Pye, Lucien. 1965 "The concept of political development." The Annals of the
American Academy of Political and Social Science Vol. 385, pp. 1-13.
Rissland, Edwina L., and Kevin Ashley. 1986 "Hypotheticals and Heuristic Device."
Proceedings from the Fifth National Conference on Artificial Intelligence, Phila-
delphia, Pennsylvania.
Ross, Brian. 1989 "Some Psychological Results on Case-Based Reasoning." Pro-
ceedings of the Case Based Reasoning Workshop. Pensacola Beach, Florida.
Rudebeck, Lars. 1970 "Political development: towards a coherent and relevant
formulation of the concept." Scandinavian Political Studies.
Sartori, Giovanni. 1975 Tower of Babe/: On the Definition and Analysis of Concepts
in the Social Sciences. Occasional Paper Number 6, International Studies As-
sociation, University of Pittsburgh.
Toulmin, Stephen. 1963 Foresight and Understanding: An Inquiry Into the Aims of
Science. New York: Harper and Row.
Waltz, Kenneth. 1959 Man, the State and War. New York: Columbia University Press.
World Development. 1983 Vol. 11, No. 10.
PART THREE
AI/IR Research:
The Discourse of Foreign Policy
12
The Expertise of the Senate
Foreign Relations Committee
G. R. Boynton
On May 17, 1987, the USS Stark was struck by an Iraqi missile, disabling
the ship, killing thirty-seven U.S. sailors, and moving reflagging Kuwaiti
tankers to center stage of U.S. politics. From the time the Stark was hit
until Congress recessed that summer, the Washington Post ran three stories
a day on the attack and the Reagan administration plan to reflag Kuwaiti
tankers and protect them with the U.S. Navy. The Senate Armed Services
Committee, the House Merchant Marine Committee, and the Senate Foreign
Relations Committee held hearings. Congress passed a resolution demanding
to be fully informed before the administration went ahead with its plan.
Legislation to require the administration to take other actions was introduced
in the House and the Senate. Protecting Kuwaiti tankers was high politics
in the summer of 1987.
The administration was eventually able to convince Congress to accept
their plan, but not without great difficulty. The difficulty was threefold. First,
there was little confidence in the administration's foreign policy at that point.
The !ran-Contra hearing was in progress that same summer, and it reminded
Congress daily that the word of the administration was not always to be
trusted. The fiasco of Lebanon was another impediment to confidence in
the administration's foreign policy planning. As Senator Kerry said, "We
went into Lebanon with the understanding that we were escorting the PLO
out. That got translated into a status quo support of the Gemayel government,
which ultimately resulted in the loss of lives of Marines." 1 We do not want
that to happen again, was the position of even the most supportive.
Second, there was a discontinuity between the public rhetoric of the
administration and the plan presented to Congress by administration spokes-
people. President Reagan and Secretary of Defense Weinberger stressed the
principle of freedom of navigation and protecting the free flow of oil. lt
was an expansive rhetoric aimed at convincing the U.S. public that great
291
292 G. R. Boynton
matters were at stake. However, the plan presented to Congress involved

protecting only eleven of the 600 ships passing through the Gulf each
month. The plan did not seem to match the rhetoric, and with Lebanon in
the foreground of their thinking, members of the committee wanted to know
which represented the administration's "real" plan.
Third, it was impossible to make a realistic assessment of the danger
in the Gulf. The damage suffered by the USS Stark forcefully called to the
attention of policy-makers the dangers of the sea war in the Gulf. But on
closer inspection the sea war did not appear as dangerous as the Stark
incident had made it seem. Only 1 percent of the tankers going through
the Gulf were attacked; little oil and few lives were being lost in the war.
However, some members of the Foreign Relations Committee wanted the
War Powers Act invoked, and to make their argument the dangers had to
be stressed. The administration was equally adamant that the War Powers
Act should not be invoked, and to make their case the danger had to be
played down. Realistic assessment was a casualty of the argument.
The resulting puzzle for the Senate Foreign Relations Committee was:
What is the administration's plan?
The task for a scholar interested in the conduct of U.S. foreign policy
is determining how the committee went about discovering the administration's
plan.
Narrative Construction as a Facet of the Expertise

of the Senate Foreign Relations Committee
In a paper comparing the Senate Agriculture Committee and the Senate
Environmental Protection Subcommittee, I suggested that the expertise of
the committees was to be found in the consistent way they framed the
issues facing them, the procedures for assessing recommendations made
to them, and the legislative provisions that were used and re-used as new
circumstances arose. 2 This chapter will examine in detail the Senate Foreign
Relations Committee's construction of a narrative account of the policy-
making process. It is one of the procedures used by the committee in
assessing the recommendations of the witnesses of the administration, hence,
one facet of their expertise and one part of the answer to how the committee
went about discovering the administration's plan.
The questions and answers in the hearing of the Foreign Relations
Committee are replete with historical references, little pieces of narrative
used to justify an argument, explain a characterization, or serve as precedent.
Senator Kerry's reference to Lebanon, quoted above, illustrates one use of
narrative. He justified his "hard questioning" on the grounds that in planning
what to do in Lebanon, the administration had not carefully thought through
the limits on its policies, and the result was most unfortunate. The narrative
The Expertise of the Senate Foreign Relations Committee 293
is a reminder of a past failure that should be avoided in this situation. 3

Narrative accounts are also used as explanation, as in the following exchange.
Senator Kassebaum. The Soviet Union is an ally of Iraq, correct?

Ambassador Armacost. They are a principal military supplier of Iraq, yes.
Senator Kassebaum. And Kuwait is also an ally of Iraq.
Ambassador Armacost. Well, the Kuwaiti situation is a little more ambiguous
in the sense that they proclaim an official neutrality. It is certainly the case
that they are a nonbelligerent. They don't provide arms to Iraq. They are not
involved in the hostilities. They did not support the initiation of the war, and
they provided assistance, primarily financial, only after Iran invaded Iraqi
territory. That is, when the tide of battle shifted, they clearly perceived the
danger to their own security from an Iranian victory, and they have provided
subventions to Iraq in recent years, as have virtually all of the other moderate
Arab countries in the gulf.
They are not, however, regarded as a belligerent, and I believe Tehran has
not even so regarded them. 4
The answer to the first question is "yes." The answer to the second question
is a "history lesson," a narrative account of Kuwaiti involvement in the
Iran-Iraq war. They have not provided arms nor have they been involved
in hostilities. They did not support the initiation of the war; they became
involved only after Iran invaded Iraq. . . . But there is no "yes" or "no."
Instead, a narrative serves as an explanation. In the Foreign Relations
Committee hearing there are sixty points in which narrative is used for one
or the other of these purposes. This is substantially greater than in the
Armed Services Committee hearing on reflagging the Kuwaiti tankers, in
which this type of historical reference occurs only sixteen times.
A number of students of international relations have called attention to
the importance of history as precedent in policy making. The focus of this
chapter is on a related but less attended to use of narrative. Running
throughout the hearing is the question of when an event occurred in
relationship to other events. Did the Kuwaitis go to the Soviet Union before
coming to us? Was there a risk assessment before the March meeting?
What were we doing in the Gulf before December 1986? In asking and
answering these and related questions, the committee members and witnesses
construct a narrative account of the events surrounding the formulation of
the administration's policy. These questions dominate the conversations of
the hearing; there are seventy-one instances of questions and answers of
this form. It seems clear that constructing a narrative account of the events
is an important procedure used by the committee in determining what the
policy is and in assessing it. This procedure is not unique to the Foreign
Relations Committee, however. Similar questions are found in the hearing
of the Armed Services Committee, though they are less prominent in that
294 G. R. Boynton
committee's hearing. Nor is the procedure unique to this issue. In 1972 the
Soviet Union bought every available bushel of grain from the United States,
which tripled the price of grain and raised fears among consumers about
increasing costs of food. In the hearing held by the Agriculture Committee
in 1973, a comparable reconstruction of the event is found. Thus, it appears
to be a standard technique of congressional policy analysis, one that has
not received sufficient attention from scholars.
Research on Narratives and Cognition

Narrative accounts of actions through time play a central role in human
communication and cognition. Waiter Fisher, a distinguished scholar of
communications, has argued for the centrality of narrative accounts in human
communication. 5 Frederick Olafson, 6 a philosopher, and Hayward Alker7 have
analyzed and specified the structure of narrative accounts. Stephen Read 8
has reviewed the cognitive science research on the operation of narratives
in cognition. Pennington and Hastie, cognitive psychologists, have shown
the importance of narrative in making judgments in a jurylike setting. 9
Working in the Schank-Abelson tradition, Dyer has done a very thorough
analysis of understanding the narratives of everyday life. 10
In studying international affairs, Hayward Alker's work on precedent logic
has been an important influence. 11 Using different tools, Dwain Mefford 12
and Phil Schrodt 13 have pushed precedent-based explanations of political
action forward. In a paper on news coverage of the Kuwaiti tanker reflagging
controversy in the Washington Post, I showed that the coverage could be
encoded in a model of memory in a way that would permit reconstruction
as a narrative account. 14
This chapter is most like the work of Pennington and Hastie in its design,
while relying on John Anderson's theory of cognition 15 for its basic theoretical
underpinning.
In their research, Pennington and Hastie showed a three-hour video
recording of a trial to subjects. After watching the video recording the
subjects were asked to talk out loud as they acted as jurors and decided
the innocence or guilt of the defendant. The protocols were then analyzed.
They found the subjects uniformly reconstructed the questions asked and
the answers of witnesses in a narrative account of the actions leading to
the arrest and trial of the defendant. They summarize this result in the
following way.
The first major empirical finding is that spontaneous interview protocols do

exhibit story structures. When subjects were asked to talk aloud about their
decisions, content coding of the protocol text yielded individual and verdict
group structures that have causal event chain and episode structures. Jurors'
stories were not simple lists of evidence. They always contained most com-
ponents of the episode schema (initiating events, psychological and physical
states, goals, actions, and consequences) in appropriate causal relations. Jurors
typically inferred missing components for these structures when they were
not contained in direct testimony at trial. Evidence not related to a story of
what happened was systematically deleted from discussion. In addition, jurors'
stories contained events and goals that were sensibly related to verdict category
features. Specifically, psychological states and goals of actors in the stories
were related to features of the verdict categories designated as legal elements
of the charge against the defendant. 16
The search for "initiating events, psychological and physical states, goals,
actions, and consequences" that is found in the text of the hearing of the
Senate Foreign Relations Committee is also found by Pennington and Hastie.
This suggests the need for a cognitive model of reconstructing a narrative
account of episodes. Anderson's theory of cognition suggests a number of
features that should characterize such a model. In Anderson's theory there
are three representations of knowledge: as spatial images, as temporal strings,
and as propositions. The latter two are important for this chapter. Both are
stored in declarative memory and are accessed or brought into working
memory through the spread of activation. The two are differentiated by the
operations that can be performed on them. Unlike propositions, components
can be inserted into and deleted from strings. Strings are front- and back-
end loaded; they have a beginning, an end, and the rest is middle, which
is parallel to a narrative. Before and next are operations that can be performed
on strings. Propositions are abstract in roughly the way a declarative sentence
is an abstraction of one aspect of an experience. There are strong constraints
among the elements of propositions. "Thus 'hit' takes two arguments, 'give'
three, and 'decide' must have as one of its arguments an embedded
proposition." 17
Anderson calls both strings and propositions cognitive units because they
are handled the same way in memory. There are limits on how much can
be encoded in a single cognitive unit. Large knowledge structures are encoded
hierarchically, with smaller cognitive units embedded in larger cognitive units.
The hierarchical structures can combine strings and propositions. In this
combination, Anderson shows that it is possible to represent the "scripts"
that Schank and Abelson posit to handle larger units of action. Anderson's
theory is a production system theory, and it is productions that perform
the operations on declarative knowledge. A production is a condition-action
pair. The condition specifies a data pattern, and if that data pattern is in
working memory the production will operate and the action will be taken.
Control of cognition is conceptualized as inherent in the productions through
the incorporation of goal statements in productions. Productions are learned
296 G. R. Boynton
by imitation and recipe and become compiled through practice. Once they
are thoroughly learned, the person may not be able to consciously reconstruct
all the steps in the production. Anderson's theory is architectonic; it is a
theory of the elements of cognition and how they work together. This
extremely brief overview mentions only those elements that are important
in this research. 1B
In Anderson's theory, research on cognition involves building a model of
the cognitive action incorporating both the productions and the declarative
knowledge involved. Any model will be data-intensive, paralleling work in
cognitive science on expert systems. 19 Abelson, from the schema side of
cognitive science, has recently made the same point.
Theories of cognition must contain theories of content. Newell has also argued
that a theory of cognition in a particular domain first demands a theory of
the domain itself, which he calls the "knowledge level." One component of
a complete model is a description of the more complex structures in memory
and their characteristic properties. Although differences between such structures
may be reducible in part to more general representational principles, such as
number of associations or level of abstraction, these properties are not sufficient
to capture the observed differences. 20
This agreement from all sides of cognitive science suggests what one should
be looking for in a theory of narrative construction. It will not be a small
number of principles. Instead, it can be expected to be an elaborate collection
of productions and hierarchically organized declarative knowledge. Strings
will provide the sequential character of the narrative, and propositions about
the United States, the Soviet Union, and the Gulf will abound.
The Committee Hearing

The Senate Foreign Relations Committee met for public hearings on May
29 and June 16 of 1987 to question witnesses from the administration about
the administration's plan to reftag and protect the Kuwaiti tankers. At the
second meeting, four other witnesses testified briefly. As is standard for
congressional committee hearings, the chairman of the committee, Senator
Pell, was the only senator who was present through the entire testimony.
These were not, however, hearings in which the chairman was the only
senator present. Members of the committee were actively involved in
questioning witnesses, and there is evidence in the record that senators
were also present when it was not their turn to ask questions.
A verbatim record of the opening statements and the questions and

answers was published. These parts of the record were put into a file for
computer-aided processing; the file contains 47,000 words. The file was
divided into conversations based on the manifest subject being discussed.
(A conversation normally begins with a question from a senator about a
subject and continues until the subject is changed.) The file includes 162
conversations; of these, seventy-one involve reconstructing the history of
the policy-making process. These seventy-one conversations are used in
this chapter.
I need to be clear about my use of committee hearings. I am not claiming
they are the principal vehicle of learning for members of the committee.
There are many places in the text where members assert that they have
learned something they did not know, but it is difficult to imagine what
evidence would be relevant to an assertion that this is their principal means
of learning. I am claiming that in at least some hearings it is possible to
learn how committee members understand what they are doing in assessing
recommendations made to them by studying their conversations with wit-
nesses. The verbatim record of these conversations is a "site" for research
on cognition.
Constructing a String of Events

For this paper, a narrative will be defined as a sequence of events for
which there is a plausible connection between each contiguous pair in the
sequence. This is a definition appropriate for an observer engaged in
reconstructing a narrative from the How of questions and answers-which
do not proceed from beginning to end of the sequence-of a committee
hearing. Many senators ask about certain parts of the sequence. One or a
few senators ask about other parts of the sequence. And the order in which
events are revealed may have little to do with the order of the calendar.
This is the situation of the juror in the Pennington-Hastie research, just as
it is the position of members of the Foreign Relations Committee as they
listen to their colleagues ask questions.
lt should be noted that this is a different cognitive task than constructing
a narrative. For any collection of historically contiguous pairs of events
there are many possible end states. The task of constructing a narrative
is devising a sequence of pairs of events that leads to a specific end state.
The observer's task is substantially simpler; it is to order pairs of events
that are given in what is being observed and detect the plausibility of the
connections.
298 G. R. Boynton
This section investigates the task of constructing the sequence of events

and then examines the sequence that is constructed. The next section
investigates the plausibility of the connections.
What Does One Need to Know to Reconstruct the Sequence?

Testimony before the Senate Foreign Relations Committee does not begin
at the "beginning" and trace events through to the end. The cognitive task
is reconstructing the historical sequence from the scattered references in
the questions and answers of committee members and witnesses. A series
of small segments of the text will illustrate the procedures required in
constructing the sequence.
The opening statement of Senator Pell illustrates what one needs to know
about the calendar to construct the sequence.
Friday, May 29, 1987

The Chairman. The Committee on Foreign Relations will come to order.
This morning, our committee will hold an open hearing on the current
situation in the Persian Gulf, the implications of the administration's plan to
put American flags on escort Kuwaiti-owned tankers, and the refusal of the
administration to abide by the requirements of the War Powers Resolution.
The tragic attack May 17 on the USS Stark has focused public attention
on the perils facing our ships and crews in the Persian Gulf. . . . 21
Last Thursday, the Senate approved by a vote of 91 to 5 an amendment
to the Supplemental Appropriations Bill requiring a report on security ar-
rangements in the Persian Gulf prior to implementation of any agreement for
United States military protection of Kuwaiti shipping. 22
"This morning," "May 17," and "last Thursday" are the clues to putting
these three events into the right order in a sequence. The various forms
of reference to the calendar will handle putting many of the events in orderP
Senator Moynihan. Mr. Chairman, the United States has been in the Gulf
since 1949. 24
The point that has to be recognized in this statement is the reference to

an ongoing activity. The United States has been-and continues to be-in
the Gulf.
Ambassador Armacost. There had been interagency discussion in January

and February. 25
There are a few cases like this in which the date is indefinite. In order to
fit it into the sequence, other events in the month have to be compared
with this one to determine the order. The president made a statement on
January 23. The interagency planning must have started before that pres-
idential statement. So the interagency discussion is placed before the
presidential speech in the string.
Calendar "arithmetic," the distinction between completed and ongoing
events, and some comparison of pairs of events are required to establish
the order of the sequence of events. Reading a list, comparing items, and
insertion are the procedures for constructing the sequence.
Analysis of the Chronology

The United States had been in the Gulf since 1949. The Soviets had
been relatively uninfluential in the Gulf for fifteen years and were seeking
ways to increase their influence. The sea war between Iran and Iraq started
in 1984. Kuwaiti tankers were targeted in September and October of 1986.
The U.S. sale of arms to Iran became public knowledge in November 1986.
The Kuwaitis turned to the Soviet Union for assistance and then requested
aid from the United States. When the United States learned that they were
to share the protection of Kuwaiti tankers with the Soviet Union, the
administration offered to protect all eleven. These arrangements were being
made with very little attention focused on them until the attack of the USS
Stark. After the attack the administration reassessed the military require-
ments. Congress and the public got into the act.
This is the chronology of events in broad outline. The sequence of events
taken from the hearing is composed of thirty-nine events. If the sequence
was arrayed on time what would be seen is a very sparse set of events
until the sea war began, and then an increasingly dense distribution ap-
proaching the summer of 1987, which ends with the second committee
hearing.
The administration also composed a chronology of events. There are
thirty-four events in their chronology. It begins with November 1, 1986, the
date Kuwait publicly expressed its concern about the attacks on its ships,
and it ends with the attack on the USS Stark. Thus, it covers a much
shorter period of time. The congressional chronology extends farther back
in time and also covers events after the attack on the USS Stark. (Also,
the administration chronology does not mention the arms sales to Iran
becoming public knowledge.)
Summary
The chronology is not yet a narrative; it is only an ordered list. The
"glue" of plausible connection is not explicitly part of the chronology. A
broad sweep of history is covered, becoming quite detailed for the period
300 G. R. Boynton
after the attack on the USS Stark and around the committee hearings. lt
is to "plausible connection" that I turn next.
Plausible Connections
This committee and other congressional committees construct narrative
accounts of the policy-making process in making their assessment of the
policy. lt is not all that is going on in this hearing, but it is a major focus. 26
Two quotations involving Senator Moynihan show the global understanding
most of the committee members brought to the hearing and introduce the
primary mechanism for establishing plausible connections.
Mr. Chairman, the United States has been in the Gulf since 1949. lt is a
region that is essential to the West and to Japan and the democratic nations
of the world. lt is not essential to the Soviet Union save that they may have
an opportunity to deny it to us. The Soviets export oil and the Gulf exports
oil.
We cannot let the Persian Gulf become a Russian lake. The Kuwaitis have
done us no great service by inviting the Russians in. We have been there,
protecting them in one way or another for a generation, and suddenly, last
November, I believe, as the testimony will bear out, they invited the Russians
in and the Russians saw that open door.
This is Scoop Jackson's 75th birthday, and I remember his description of
the Soviets as "acting like hotel thieves." They just go down a hallway and
test every door until they find one that is open. They found one that was
open, and I think we have to assess that situation. 27
In this brief statement, Senator Moynihan gives his interpretation of the

importance of the Gulf to the United States and Western democracies. The
Gulf is important because it is a major source of energy. He also gives his
interpretation of the importance of the Gulf to the Soviet Union. They keep
looking until they find a door that is open, and when they find an open
door they enter to create problems for the West. But there is one puzzle
to which he has no answer. The United States has been in the Gulf protecting
the Kuwaitis. Why would the Kuwaitis let the Russians in? lt is a pair of
events for which he has no plausible connection.
Later, while questioning the administration witnesses, Senator Moynihan
is given a connection he finds plausible.
Senator Moynihan. But, sir, it cannot have given you any comfort at all.
They have that tie. We have protected them. We have let those billionaires
grow to be multi-billionaires, and the minute it is to their convenience, they
let the Soviets in.
Ambassador Murphy. Well, the moment of that decision was not coincidental.
lt was November 1986.
Senator Moynihan. lt was not coincidental that the Kuwaitis asked the
Soviets in as it came to be known through the world that we had been
shipping arms to Iran?
Ambassador Murphy. lt's got to have been a factor. . . .
Senator Moynihan. Mr. Chairman, I thin~ Ambassador Murphy has said
something very profound. lt was the aftermath of the revelation of the deceptive,
dishonest policy of shipping arms to Iran that some persons in the administration
carried out, that Kuwait turned to the Soviet Union for protection. 28
It is the same question. Look at how we have protected them. How could
they let the Soviets in "the minute it is to their convenience?" The answer
is, that was the time the lran-Contra sale of arms to Iran became public
knowledge. Senator Moynihan had found what I will call an interpretive
triple: two events with a plausible connection. The two events are U.S. sale
of arms to Iran and Kuwait turning to the Soviets for assistance. The
interpretation is: Given what we had done, the Kuwaitis did not know if
they could trust us to take actions that were beneficial to them any longer.
The Kuwaitis had difficulty construing the sale of arms to Iran, whatever
the reasons of the Reagan administration, as actions benefiting them. Hence,
protect yourself by turning to the other superpower. This made sense to
Senator Moynihan. It also appeared to make sense to many of the members
of the committee because it became a lynchpin in their analysis of the
history of the development of the policy.
Interpretive Triples
Forty interpretive triples are found in the hearings. However, I will show
that interpretive triples are used for two purposes in addition to constructing
the narrative. In this section, a number of interpretive triples will be
summarized to illustrate the style of argument.
Senator Boschwitz's is the most succinct of the interpretive triples. "I
certainly agree that the Kuwaitis probably play us like a violin. They go to
the Russians and then they turn to us. " 29 Action 1: the Kuwaitis went to
the Russians. Action 2: they came to us. Interpretation: they played us like
a violin. This interpretation draws on exactly the same goals for the two
countries that Senator Moynihan articulated. The Soviets want in; we want
them out. But it turns those motives into a strategy for playing the
superpowers. The action of the Kuwaitis makes sense, though it is a
somewhat different sense than Senator Moynihan made of it. While Senator
Boschwitz has the most succinct version of this interpretive triple, it occurs
over and over as the members of the committee search for the "real"
reason for protecting the Kuwaiti tankers.
302 G. R. Boynton
Senator Pell opens the first meeting of the committee with an interpretive
triple. Action 1: the attack on the USS Stark. Action 2: the committee
hearing. Interpretation: it "focused public attention on the perils facing our
ships and crews in the Persian Gulf. "30 We meet to consider our policy in
the Gulf because the attack on the Stark has focused public attention on
the perils in the Gulf.
Administration witnesses use an interpretive triple to show that one of
the actions is implausibly connected to the other.
Ambassador Murphy. Yes, Senator. But I refer back to hearings, closed

hearings before this committee, and the House Foreign Affairs Committee,
briefings provided to both individual members of Congress and to staffers
when other committees were not available, in March, about the effort and the
thought we were giving to this program of refiagging.
We did not try to hide it. In fact, we came up here very deliberately and
very quickly, when the President had made his decision in March.
So, to the extent that we stand accused of not consulting and not being
ready to consult with the Congress, I regard that as a "bum rap."
Action 2: congressional criticism that the administration is not willing to

consult with congress. Action 1: administration attempts to consult. Inter-
pretation: it is "a bum rap." This is a triple that occurs several times in
the hearing.
Interpretive Triples as Counterfactual Arguments

An interpretive triple can be used to ask a counterfactual question by
suggesting an action, an interpretation, and an action that should have
followed but did not. Why not? is the question. Senator Kassebaum asks
a question based on this kind of counterfactual.
Senator Kassebaum. These objectives [freedom of navigation and free flow

of oil] were ones that were just as valid, it seems to me, in 1986 as they are
in 1987.
Ambassador Murphy. Yes.
Senator Kassebaum. Did we have to wait until the Kuwaitis asked us?
believe you said the Kuwaitis asked us.
Ambassador Murphy. On those objectives, Senator . . .
Senator Kassebaum. But what I meant by the objectives was more in line
with the shipping decision and if we had calculated previous to the January
decision that we should get more involved.
Ambassador Murphy. We didn't have a request until December. That was
when the request came to us from Kuwaiti authorities. 31
Action 1: disruption of shipping in the Gulf in 1986. Interpretation: we are

doing this to maintain freedom of navigation and free flow of oil. Counterfactual
Action 2: why not greater involvement in 1986? Senator Kassebaum takes
the rationale for the current action, pushes it back one year, and asks why
it was not equally appropriate for taking action then. The only answer is
we did not have a request until December. That, of course, is no answer
at all if the rationale is the reason for taking action, Senator Kassebaum is
pointing out.
Interpretive Triples as Predictions

Another use of interpretive triples is making predictions about what will
happen if the U.S. Navy does or does not protect the reflagged Kuwaiti
tankers as proposed by the administration. This is a form of pushing the
narrative into the future. It is particularly important for the argument about
the applicability of the War Powers Act, but it is used for making other
predictions as well. There are seventeen instances of interpretive triples
being used to consider what might happen in the future.
If the United States refused to reflag the Kuwaiti tankers and protect
them, would the Russians provide protection to the Kuwaitis? The answer
takes the form of a predictive interpretive triple. Action 1: when asked in
November to protect Kuwaiti tankers, they agreed. Interpretation: the same
reasons that were applicable then would be applicable if we refused to
protect Kuwaiti tankers; they still want to get in. Predicted Action 2: Yes,
they would offer protection. This, at least, is the answer of the administration.
Prediction is central to assessing the risk of the administration's policy,
and the administration witnesses were asked about it numerous times. Their
answer was always the same. Action 1: in the sea war for the last three
years the Iranians have been very cautious when ships of the U.S. Navy
and other navies have been around; no U.S. flagged ships have been attacked.
Interpretation: all we are doing is adding a limited number of U.S. flagged
ships to be protected. Predicted Action 2: we expect, though we certainly
cannot guarantee, that Iran will be as cautious in the future as they have
been in the past, and no U.S. flagged ships will be attacked.
This argument is attacked in several ways. The attack on the USS Stark
was cited. The increasing attacks on shipping during the three years of the
sea war was used. But the interesting arguments involve shifting the
interpretive triple. Senator Cranston makes that shift in one of the sequences
of questions and answers about risk.
Senator Cranston. Was that based upon the administration's analysis of how
the Ayatollah Khomeini, who is a rather unpredictable leader, to put it mildly,
304 G. R. Boynton
would behave in response to what amounts to the United States, viewed by

him as the "Great Satan," intervening on behalf of Jraq?32
Senator Cranston substitutes Ayatollah Khomeini for the record of the three
previous years. Khomeini's statements about the United States as the "Great
Satan" and unpredictability produces an interpretive triple with some plau-
sibility in which Action 2 is attack on U.S. ships. When that does not
change the administration answer, he tries one more challenge.
Senator Cranston. We have not been escorting ships up to now that have
been helping Iraq.
Mr. Armitage. We have not been escorting ships that have been helping
Iraq.
Senator Cranston. So the assumption that we are making is based upon
circumstances quite different from the circumstances that would develop if
we start escorting ships carrying oil from lraq. 33
He questions the relevance of the first action by noting that U.S. ships in
1984-1986 had not been helping Iraq in any way, but if Kuwaiti tankers
are protected that will no longer be true. Kuwaiti income derived from selling
oil is used to pay part of Iraq's war expenses. That involves U.S. protection
in the war, and makes the record of the past three years irrelevant, he
believes.
Other Means of Establishing Plausible Connections

In the interpretive triples that have been reviewed, the interpretation was
explicitly made in the text. A reader can cull from what was said two
actions and an explicit interpretation. That is not, however, the only way
an interpretive triple can be articulated. lt would be difficult to read the
following statement by Ambassador Armacost as anything other than an
interpretive triple.
Ambassador Armacost. The issue for us was crystallized by a specific

request from a friendly country that had been targeted for harassment and
intimidation and which was offering or was approaching the Soviet Union as
well as ourselves for protective arrangements. 34
The language used to characterize what Iran did to Kuwait, "targeted for
harassment and intimidation," carries the interpretation. Why did Kuwait
turn to other nations for assistance? Because Iran was singling them out
for "harassment and intimidation." If the response to harassment and
intimidation was "hit me again," we would find harassment and intimidation
an unusual and probably inappropriate characterization of the action. The

plausible connection is made in the language used by the narrator in
describing events. The language shared by the narrator and the audience
carries the plausible connection. This is probably as frequent as is the
explicit statement of an interpretation. The consequence for the scholar is
the necessity of much more careful attention to the language used by the
political actors, research not widely practiced by political scientists even
though such language is one of the principal diplomatic arts.
Plausible connections are also established by embedding the action in a
well-established pattern of activity. This occurs in the hearing held in May
when members of the committee ask what the administration has done
about obtaining assistance from the Gulf states, England and France, Japan,
and a number of other actions they expect the administration to take. The
answer is always the same. There is a meeting in the White House today
with all the agencies concerned. Once the president has decided exactly
what we need then we will contact them. The planning process has gone
forward, and now it is up to the president to make the final decisions. This
pattern of activity-organizational planning ending with a decision by the
president-is well known to the members of the committee. It establishes
a plausible connection between what the administration has been doing and
future actions that will be the answer to their questions. It does not answer
the questions, of course.
Conclusion
As I understand our science, it is about explaining why the United States
sent thirty ships to the Gulf, at a cost of about $500 million, 35 to protect
eleven Kuwaiti tankers. We assume there is enough redundancy in human
affairs that what we have learned in studying other political events will be
helpful in this explanation-and that something learned in explaining this
episode will help somewhere else. In this conclusion I need to say something
about the particularity of why the navy was sent and the generality of what
can be carried from this explanation to others.
Sending the Navy to the Gulf

The Senate Foreign Relations Committee, which might have actively
opposed the plan of the administration, did not. There is no single explanation
for their action, but one part of the explanation has been examined in this
paper. They did not oppose the plan of the administration because they
could reconstruct a plausible narrative of the actions leading to reflagging
the Kuwaiti tankers.
306 G. R. Boynton
The story begins in 1949 with the initial U.S. commitment to protect
the interest of Western industrial nations in the oil of the region by protecting
the Gulf states. Eight presidents have pursued this policy, one witness noted,
and opposing the plan of the administration would be to turn one's back
on this history. lt cannot become a Soviet lake, Senator Moynihan asserted,
and other members of the committee agreed. They did not like being played
like a violin, but they could understand why Kuwait wanted protection from
Iranian harassment and intimidation. They could even understand why
revelation of the secret arms sales to Iran would lead Kuwait to turn to the
Russians before calling for the assistance of the United States. They learned
that while they had been attending to other matters, the administration had
engineered the switch that reduced the role of the Russians from co-equals
to a minor role in protecting Kuwaiti oil transport. Credibility of the U.S.
presence in the region had to be reestablished, and protecting Kuwaiti
tankers was both vehicle for and the cost of reestablishing that credibility.
With the attack on the USS Stark the administration had substantially
revised their estimate of the military presence needed and of the assistance
they would need from the Gulf states and others in carrying out the mission.
The risk involved was consistently underestimated by the administration,
much to the consternation of senators who wanted to invoke the War Powers
Act. Despite the disagreement over the risk involved, members of the
committee could not imagine a narrative ending any way other than with
U.S. ships protecting reflagged Kuwaiti tankers. All the pieces fit together,
and they all pointed to the U.S. Navy expanding its role in the Gulf.
Adversarial, sometimes angry, sometimes insulting questioning pulled out
the bits and pieces of the narrative. In its reconstruction it was a story
they could not deny.
Narrative Reconstruction
A major finding of the research is that narrative is built pairwise. The
interpretive triple composed of two events connected by a plausible inter-
pretation is the structure of much of the talk involved in reconstructing
the narrative. Only three times in the questions and answers of the hearing
does a witness or committee member string together a longer set of events.
This is completely consistent with what is known about the bounded rationality
of human cognition. The focus of attention is always narrow; elaborate
thinking is sequential in structure.
The interpretive triple is a very flexible tool of analysis. It can be used
to reconstruct an elaborate narrative-one pair of events at a time. lt can
be used for counterfactual thinking. lt can be used for making predictions.
The flexibility of the tool makes it ideal for the many facets of policy
analysis, which is undoubtedly why it is so prominent in this hearing.
Interpretive triples readily lend themselves to being rewritten as productions

or rules. It is a perfectly straightforward task to do the rewriting and, once
it is done, reproduce the outlines of the text used in the analysis. Dyer
does something much like this in his work on understanding narratives. 36
I have not done that in this paper because it understates the organization
of the thinking of members of the committee that I believe is there. The
resulting collection of fewer than forty productions would be floating in
isolation with no organization. I do not believe that reflects the cognitive
structure of the thinking of the committee. And that points to the next
step in the research: to develop a model of the understanding of international
affairs into which these interpretive triples fit coherently.
Artificial Intelligence and Research

on International Relations
The language of international relations is, for the most part, the language
of interests-not the interests of individuals, but the interests of a corporate
entity called the nation. National interest seems very distant from com-
munication, cognition, learning, and other subjects studied by the artificial
intelligence community. If one starts from artificial intelligence, however,
the bridge is not difficult to build.
A fundamental point in Anderson's theory of cognition is the explicit
emphasis on learning. In his theory there are productions and declarative
knowledge, there is representation of declarative knowledge, memory man-
agement, and control of cognition, and that is all. All ideas and actions are
built out of these basic components. All cognitive activity is productions
acting on declarative knowledge, and productions are learned just as is
declarative knowledge. A straightforward implication is that interests are
learned.
National interest may be said to be learned in the communication of one
part of the nation attempting to convince another part of the nation about
what should be done in some situation-in the questions and answers of
the Foreign Relations Committee hearing administration witnesses and sen-
ators persuade each other, and themselves, that in this situation the national
interest is. . . . The national interest is learned as the president speaks to
the nation and listens to his mail, the polls, and the politicians who advise
him. The national interest is learned as we argue with each other about
the fading Soviet menace or Third World debt relief. Values, strategy, the
geography of the region of the world that is the focus of attention, the
history of alliances and enmities are all learned in learning the national
interest. And that puts the study of communication and cognition, the stuff
of artificial intelligence, right at the center of understanding international
affairs.
308 G. R. Boynton
Notes
1. U.S. Policy in the Persian Gulf, Hearings Before the Committee on Foreign
Relations, United States Senate, Washington, U.S. Government Printing Office, 1988,
p. 85.
2. G. R. Boynton, "Language and Understanding in Conversations about Politics,"
paper presented at the 1989 meeting of the .Midwest Political Science Association.
3. This use of narrative is also prevalent in the hearings of the Senate Agriculture
Committee. See G. R. Boynton, "Telling a Good Story: .Models of Argument; .Models
of Understanding in the Senate Agriculture Committee," in Argument and Critical
Practices, edited by Joseph W. Wenzel, Annandale, Virginia: Speech Communications
Association, 1987.
4. Hearings, p. 88.
5. One of the most recent of his papers is Waiter R. Fisher, "Technical Logic,
Rhetorical Logic and Narrative Rationality," Argumentation, 1 (1987): 3-21.
6. Frederick A. Olafson, The Dialectic of Action, University of Chicago Press,
Chicago, 1979.
7. Hayward R. Alker, "Historical Argumentation and Statistical Inference: Towards
.More Appropriated Logics for Historical Research," Historical Methods, vol. 17, no.
3 (Summer, 1984): 164-173.
8. Stephen J. Read, "Constructing Causal Scenarios: A Knowledge Structure
Approach to Causal Reasoning," Journal of Personality and Social Psychology, 52
(1987): 288-302.
9. Nancy Pennington and Reid Hastie, "Evidence Evaluation in Complex Decision
.Making," Journal of Personality and Social Psychology, 51 (1986): 242-258.
10. .Michael George Dyer, In-Depth Understanding; A Computer Model of Integrated
Processing For Narrative Comprehension, Cambridge, .Massachusetts, .MIT Press,
1983.
11. Hayward R. Alker, Jr., and C. Christensen, "From Causal .Modeling to Artificial
Intelligence: The Evolving UN Peace-Keeping Simulation," in Experimentation and
Simulation in Political Science, edited by J.A. LaPonce and P. Smoker, Toronto:
University of Toronto Press, 1972, pp. 177-224; and Hayward R. Alker, Jr., and
William Greenberg, "On Simulating Collective Security Regime Alternatives," in
Thought and Action in Foreign Policy. edited by .Matthew Bonham and .Michael
Shapiro, Basel: Birkhauser Verlag, 1976, pp. 263-306.
12. Dwain .Mefford, "Formulating Foreign Policy on the Basis of Historical Analogies:
An Application of Developments in Artificial Intelligence," paper presented at the
1984 meeting of the International Studies Association.
13. Philip A. Schrodt, "Adaptive Precedent-Based Logic and Rational Choice: A
Comparison of Two Approaches to the .Modeling of International Behavior," in
Dynamic Models of International Conflict, edited by Urs Luterbacher and .Michael
Ward, Boulder, Colorado: Lynne Rienner Publishers, 1985, pp. 373-400.
14. G. R. Boynton, "Communication and Cognition in The Presentation of
International Affairs," unpublished paper, 1988.
15. John R. Anderson, The Architecture of Cognition, Harvard University Press,
Cambridge, .Massachusetts, 1983.
16. Ibid., pp. 252-53.

17. Ibid., p. 70.
18. A somewhat more elaborate description, though still brief, is found in G. R.
Boynton, "Ideas and Action; A Cognitive Model of the Senate Agriculture Committee,"
Political Behavior, forthcoming 1989. The best summary, of course, is his book.
19. Edward Feigenbaum, "Knowledge Engineering: The Applied Side," in Intelligent
Systems, edited by J.E. Hayes and D. Michie, Chichester, England: Ellis Horwood,
1983.
20. James A. Galambos, Robert P. Abelson, and John B. Black, Knowledge
Structures, Lawrence Erlbaum Associates, Hillsdale, New Jersey, 1986, pp. 16-17.
21. Hearings, p. 1.
22. Ibid., p. 2.
23. In this research I do not use a parser that can convert characters into
meaningful words and sentences. Instead, I assume meaning, which means I serve
as the parser converting the text into a stylized language that can be matched by
the program.
24. Hearings, p. 2.
25. Ibid., p. 86.
26. Senator Pell, with some assistance from Senator Cranston, made an argument
for bringing in a UN peacekeeping force. Senators Adams and Kerry argued quite
vociferously that the War Powers Act should be invoked.
27. Hearings, pp. 2-3.
28. Ibid., pp. 31-32.
29. Ibid., p. 74.
30. Ibid., p. 1.
31. Ibid., p. 22.
32. Ibid., p. 80.
33. Ibid.
34. Ibid., p. 87.
35. This is a very rough estimate. In the April21, 1988, Washington Post, Admiral
Crowe estimated it was costing about $1 million a day to protect the Kuwaiti tankers.
The protection lasted something over a year. In addition, there was the cost of
repairing U.S. Navy ships damaged in action in the Gulf and the cost of reparations
for the families of persons killed by the USS Vincennes.
36. Dyer, In-Depth Understanding.
13
Reproduction of Perception and
Decision in the Early Cold War
Sanjoy Banerjee
Robert Keohane has noted the contention of rationalist and reftectivist

approaches to the study of international relations. 1 The reftectivist approach
focuses on the interplay between the subjective concepts and discourse of
agents and social order. There is a paucity of reftectivist formal models of
international relationships, and this chapter is offered as a step in filling
that gap.
This chapter describes a computer program simulating the self-reproducing
processes of perception and decision of the superpowers in the early Cold
War. The program has two parts: a general multiagent social interaction
model and a historical data base describing some of the superpowers'
subjective concepts and material conditions of the early Cold War. Each
superpower is reconstructed as a group subject constituted by a shared
system of meaning. The program serves as a proof that the postulated
psychocultural and material conditions constituted a self-reproducing his-
torical structure. 2 The account of the conceptual system is supported by
an examination of the Soviet and U.S. foreign policy discourses of the
period.
A comparison between the present simulation and the rational choice
framework is in order. Rationalist decision making is preference-guided choice
from a menu of alternatives. Both the preferences and the alternatives are
represented with concepts not derived from the discourse of the agents. In
the present simulation, action decisions flow from subjectively encoded,
experience-dependent rules applying themselves to perceived situations.
Agents here don't really make choices at all. Rather, their actions are
controlled by the decentralized logic of their decision rules. The psychocultural
structure of an agent contains no cognitive center that can be said to make
choices. Agents are cast as intelligent but not "rational." Their intelligence
lies in the conditions of demise and survival of such rules of action.
310
Reproduction of Perception in the Early Cold War 311
The broader differences between rationalist models and this work lie in
the significance they attribute to the social heterogeneity of thought and
perception in stable group interaction. Rationalists freely acknowledge that
persons and groups have distinctive forms of thought. They contend, however,
that a rational form of thought exists and that idiosyncratic variations from
rationality do not matter in the long run. Over time, actors return to rationality
and an equilibrium of interaction results. In this simulation, distinctive forms
of thought of interacting, communicating agents feed on each other. The
subculture of each interacting group is held to be influenced by the other
groups' subcultures, all of which cohere within a single integrated culture
and a single historical structure. The idiosyncracies of thought are seen not
as random aberrations but as the definitive meanings of the culture. The
scientific puzzle addressed here is to reveal the structures and processes
of integration, and thereby of disintegration, among meanings and practices
in a historical structure.
Rationalists have cited the "as if" criterion in rejecting the criticism that
their assumptions about cognitive processes are unrealistic. Waltz, 3 Achen
and Snidal, 4 and others have allowed that rationalist assumptions on these
matters may be unrealistic. They assert, however, that predictive accuracy
is the sole criteria for judgment of scientific adequacy. In the face of this
rationalist defense, demonstrations of the inaccuracy of rationalist assumptions
relating to mental processes are ineffective in challenging the validity of
their approach. The "as if" defense, though, has a point of vulnerability.
Relying exclusively on the claim of predictive accuracy, this defense is called
upon to give an account of its empirical methodology. Since the predictions
of rationalist theories focus on intended actions, their exponents must justify
their method of coding actions.
The central flaw in the "as if" approach lies in the relationship between
imposed categories of intentional action and their putative empirical referents.
Jervis5 has observed that the coding of historical actions as cooperation
or defection by prisoner's dilemma modelers has sometimes been arbitrary
or inconsistent with the actors' own conception of their actions. Rationalists
must use theoretical typologies of intentional actions not explicitly related
to the subjective concepts of the actors. And they must code historical
actions according to these typologies. The absence of explicit connections
between theoretical categories and actors' subjective concepts introduces
an element of arbitrariness into such a classification.
Rational choice theories often classify possible actions according to
categories that refer to actors' intentions. Intentions exist as systems of
concepts in the minds and communications of the actors, however. Any
description of intentions that makes no explicit reference to those subjective
concepts must either rely on the reader to do so or remain unverifiable. In
312 Sanjoy Banerjee
either case, a rigorous and consistent scheme of classifying intended actions

is not achieved.
The reliability of positivist schemes of classifying intentional action is of
great importance to rationalist claims of scientific accuracy. For to the
extent that the classification scheme is arbitrary, it fails to protect against
the categorization of actions with a bias in favor of the theory being tested.
This leaves the method open to inflated claims of predictive accuracy. Since
the rational choice approach relies exclusively on its predictive accuracy
for scientific validity, any inflation in the claims of such accuracy propor-
tionately diminishes the scientific validity of the whole approach.
The Reproduction Model

The reproduction model describes the repeated cognition and interaction
of multiple subjects, focusing on the processes of perception and decision
making. Each subject is held to perceive social activity at two levels: actions
and acts. 6 Subjects distinguish actions in concrete, nearly physical terms.
Acts are represented in more socially meaningful and emotionally charged
terms. Acts are interrelated and combined into scripts. The process of social
perception is cast in the simulation as taking concrete actions and their
historical contexts and matching them to acts and scripts from the subject's
cultural memory. The decision process is cast here as flowing from the
subject's perception of its situation and producing the acts/actions to be
committed.
The model shows how a historical structure can reproduce itself. It traces
the process by which each round of actions is perceived by each subject
according to its own script, and by which that perception leads to the
reiteration of a pattern of action in the next round that is again perceived
by the subjects as fulfilling their respective scripts. The early Cold War is
shown to conform to such a process.
Perception
The process of social perception is reflective. What is perceived at one
time depends on what was perceived earlier. Perception entails attaching
socially meaningful categories of acts to observed concrete actions on the
basis of the social context in which the actions are observed. A key issue
in modeling social perception, then, is how to cast the construction of
contexts by subjects. Subjects construct historical contexts in terms of
clusters of interrelated acts called scripts. Current actions are construed as
acts based on how they relate to the script perceived the previous round.
Acts and Actions. Both ordinary and legal language make an incipient
act/ action distinction. Deciding whether what someone did was a "hostile
act" or a "criminal act" is a common problem. For example, a state's

action of sending its troops across an international boundary to fight alongside
local government troops against a guerrilla army may be construed by
various perceivers as an act of imperialist aggression or an act of aiding
a beleaguered ally. Action-descriptions are assumed to be shared by all
interacting subjects in a historical structure. Act-descriptions are held to
be unique to each subject. People who disagree on what act was performed
can often agree on a concrete description of the action.
The matching of actions to acts is illustrated by the following example.
The U.S. action of refusing Soviet demands in the Marshall Plan negotiations
in the fall of 1947, in conjunction with its continuing nuclear buildup, was
construed in Moscow and in the 1947 Cominform meeting as an act of
pursuing world domination by the formation of a military and economic
capitalist bloc in Western Europe. Andrei Zhdanov, member of the Politburo,
was the chief speaker at the Cominform meeting. He argued that the United
States was capturing the position of the leading imperialist power at the
expense of the European imperialist countries, including Britain. He asserted
"the United States proclaimed a new frankly predatory and expansionist
course. " 7 Zhdanov added,
The expansionist foreign policy inspired and conducted by the American

reactionaries envisages simultaneous action along three lines: (a) strategical
military measures, (b) economic expansion, and (c) ideological struggle. 8
Short-term consequences of this interpretation were the termination of the

"people's democracy" period in Eastern Europe, including the Czech coup
of February 1948, and a wave of strikes by Communist tabor unions in
Western Europe.
Since actions are described as unique historical events, the account of
actions in a historical structure must read as a chronology. In this simulation,
actions occur in rounds. Each round of actions represents a brief historical
period. The rounds are chosen to represent six- to twelve-month periods
of European history during 1945-1948. lt is assumed that actions within a
single round are contemporaneous, and not reactions to each other. A listing
of the rounds of action in the simulation is given in Figure 13.1.
Descriptions of actions and acts are distinguishable objects in historical
discourses. Action-descriptions are those referring to behaviors unique in
time and history, whereas act-descriptions refer to repeatable instances of
behavior categories. Yet acts are not types of actions. A single action may
be a constituent in more than one act in the interpretation of a subject.
An act may comprise several simultaneous actions. The categories and
language of act-descriptions are private to subjects; the categories and
language of action-descriptions are widely shared in an historical structure.
314 Sanjoy Banerjee
Figure 13.1: Rounds of Action
Round 1:
US: build atomic bombs (BAB), promote US-controlled relief (PUR),
no atomic deployment (NAD)
USSR: promote Polish communists (PPC), allow noncommunists (ANC)
Round 2:
US: build atomic bombs (BAB), no atomic deployment (NAD), trade
expansion (TE)
USSR: promote peoples' democracies (PPD), allow Czech democracy
(ACD)
Round 3:
US: build atomic bombs (BAB), no atomic deployment (NAD), Truman
doctrine (TO), propose Marshal! Plan (PMP)
USSR: noncommunists continue (NCC), expand communist role (ECR)
Round 4:
US: build atomic bombs (BAB), no atomic deployment (NAD), promote
Western European negotiations (PWEN)
USSR: military restraint (MR), Promote Czech coup (PCC)
Round 5:
US: build atomic bombs (BAB), no atomic deployment (NAD), fund
Marshal! Plan (FMP), unify West Germany (~"WG)
USSR: military restraint (MR), consolidate communist parties
(CCP)
Act-categories are evaluative, motivating, and socially reflective. Act-

descriptions-aggression, humanitarian aid, defending the homeland-are
connected to the deepest motives of the subjects who hold these categories.
They shape the subject's emotional orientation toward the self and others.
Act-categories implicate the basic character of the actor in the eyes of the
subject. A country that commits the act of aggression is expected to do
certain other things as well. Action-categories are none of these things.
The Soviets undertook the actions of supporting mixed governments in
Eastern Europe, with Communists controlling key, but not all, ministries
during 1945-1947. These actions were interpreted by the U.S. Democratic
and Republican leaderships as acts of totalitarian aggression. President
Truman's 1947 State of the Union address-the Truman Doctrine speech-
asserted that the world was being divided into two. The U.S. leaders came
to suspect Soviet ambitions to control Western Europe as well. Powerful
emotions of anger, fear, and hatred were generated in the U.S. polity as a
result of this social construction of the act.
The Relation Between Acts and Actions. In the representation of actions
employed, each action is viewed as an intended reaction to prior actions
by the self or others. These intentional reaction relations between actions
are defined here at a very concrete level. lt is assumed that belligerents in
a conflict would agree on what the concrete intentional relations among the
actions in their conflict were; their differences would emerge at a higher

level of abstraction. Thus, the existence of a minimum perceptual consensus
in all enduring historical structures is assumed, and that consensus is the
base for conceptual differentiation among the interacting subjects. If one
troop movement action is intended to offset another movement by another
actor, then "offset" is the intentional relation between the two actions. If
the action of conditionally offering economic aid is intended to forestall
another action, the two are related by "conditioned to forestall." Since each
action is an intended reaction to some actions in the previous round, a
matrix of intentional relations among actions in each pair of consecutive
rounds can be generated.
The action/act relation is illustrated by examining the first two rounds
in the simulation. The first round represents late 1945, and the second
round, 1946. In late 1945, the United States developed its atomic weapons
without Soviet consultation or participation; this is called the action "build
atomic bombs." The United States cut off funding for the United Nations
Relief and Reconstruction Association in 1945. This had been a relief agency
that distributed relief funds in Europe outside U.S. political control. New
relief agencies were funded under Washington's political control (promote
U.S.-controlled relief). The USSR continued to support the Communist
government of Poland and placed Communist parties in powerful positions
in the other Eastern European countries in its military zone (promote Polish
Communists). Moscow allowed noncommunist groups to participate in
government at this time (allow noncommunists).
In the second round (1946), the United States proposed the Baruch Plan,
which would enable the United States to retain its monopoly on nuclear
technology but would bring the disposition of nuclear weapons under
international supervision with Soviet participation (propose Baruch Plan). In
spite of having some atomic weapons, Washington deployed none in Europe
(no atomic deployment). The United States sought to expand capitalist trade
between itself and Western Europe and within Western Europe (expand
trade). The USSR developed in Eastern Europe several governments with
Communists in key ministries but with noncommunists in high positions
as well; these were called people's democracies (promote people's democ-
racies). In Czechoslovakia, Moscow refrained from interfering substantially
in the workings of a liberal democratic government (allow Czech democracy).
Actions in the second round were intended as reactions to those in the
first. For example, the U.S. action of trade expansion was intended to show
the Europeans the superiority of the U.S.-Ied capitalist system over the
Soviet-led system developing in the East (differentiate). The Soviet action
of allowing Czech democracy deliberately underutilized the potential for
control created in its prior action of promoting Communist forces in its
zone of control (underutilize). The U.S. nuclear weapons program was intended
316 Sanjoy Banerjee
Figure 13.2: Matrix of Intentional Relations of Round 2 to Round 1

Round 1:
- BAB PUR NAD PPC ANC
Round 2:
TE gain differentiate
support
BAB guard against
NAD under-
utilize
PPD against against gain
support
ACD under-
utilize
to guard against any Soviet military threat to its earlier reconstruction

activities (guard). The Soviet actions of establishing people's democracies
was intended to build a structure of local political support for its previous
promotion of Communists in Poland and throughout Eastern Europe (gain
support). The people's democracies were a counter to U.S. actions of building
nuclear military power and reconstruction efforts under its control (against).
A matrix of intentional relations between the actions of any two consecutive
rounds emerges by this method. Those matrices represent shared beliefs
of interacting subjects. The first and second rounds are related by the
intentional relations matrix in Figure 13.2. (The matrix is read as indicating
the relation of the action in Round 2 to the indicated action in Round 1.)
Intentional relations matrices for each pair of consecutive rounds are ex-
ogenously specified in the historical data base.
Acts are subjectively constructed in terms of their constituent actions'
intentional relations to actions/acts perceived earlier by the same subject.
The U.S. act of "defending the West" is defined in the simulation as
"underutilizing" the military potential created in the prior act of "defending
the West," "guarding" its earlier act of "build a free society" (in Western
Europe), and "against" the Soviet act of "totalitarian aggression." Under
this definition, a set of U.S. actions are construed by the United States as
the act of "defending the West" if they, in any combination, hold the required
relations to actions in the previous round construed to constitute the required
acts. Thus, how actions are construed as acts in one round depends on
how actions were construed as acts in the previous round. The construal
of actions as acts is reflective; it is a dynamic system whose current
performance depends on its recording of past performance. A temporary
historical continuity of perception can be maintained this way. Causes of
breakdown of such a reflective process will be discussed later.
The United States' third-round action of "propose Marshall Plan" is

constructed as a constituent in the act of "defend the West." This act is
"against" Soviet "totalitarian aggression." Such a construction is discernible
from discourses. Senator Arthur Vandenberg said, in defending the proposed
plan from a conservative Detroit columnist's attack,
I respectfully submit we do "know enough" to know what will happen if it,

or something like it, doesn't work. We know that independent governments,
whatever their character otherwise, will disappear from Western Europe; that
aggressive communism will be spurred throughout the world. 9
Scripts. Act-concepts are parts of larger discourses constructing and

expressing scripts. Clustered patterns of acts, recognized repeatedly in the
historical experience of the subject, are scripts. Through scripts, subjects
are able to impose coherence on a complex situation. lt is the connections
among acts in a script that make acts meaningful to the subject. Indeed,
as we shall see, recognition of an act often depends on the recognition of
the whole script containing it.
The Soviet Cold War script in the late 1940s is represented here as a
cluster of five acts. Soviet-aided economic reconstruction along socialist
lines in Eastern Europe was conceived as "building socialism." The estab-
lishment of Leninist political structures in the region was part of the act
of "defending socialism" from hostile elements within Eastern European
societies who were allied with the West. U.S. economic policies were seen
as "rebuilding capitalism" after the cataclysms of fascism and war resulting
from the contradictions of capitalism. U.S. military and economic efforts
to limit the sphere of socialism and expand that of capitalism were cast
as the quest for "world domination." And elements of restraint in U.S. policy
were seen not as natural peacefulness but as "prudent restraint" due to
respect for Soviet strength.
Indeed, each act is defined in terms of its intentional relations to other
acts in its script. Thus, for an act to be recognized, much of its script
must have been recognized in the previous round. The recognition of acts
is a reflective process. Further, the intentional relations among acts in a
script form a matrix. The Soviet script matrix for the simulation is given
in Figure 13.3.
After the new Soviet script was first applied to superpower activities, its
internal logic of intentional relations served to perpetuate construal of the
whole script. For example, the Soviet Communists believed that socialism
and capitalism were the two contending systems to guide the world future,
and that capitalism was still the more dominant. Actions by the United
States designed to expand its dominance even further and limit the sphere
of socialism were regarded as a quest for total world domination. Thus,
318 Sanjoy Banerjee
Figure 13.3: Soviet Cold War Script

---~------ - ----------- -
Previous round:
r---
build defend build world prudent
socialism socialism capitalism dominatio n restraint
- - +------- -------·- -----
------l------~
Round:
build gain. 1against
social ism support II
defend f.Juard under- against

social ism utilize
build gain
capita lism support
pursue against against
world plan
dominat ion
prudent under-
restra1 nt utilize
---
once some U.S. activities were marked as a quest for world domination, all
Soviet actions designed to thwart U.S. actions so marked were acts of
"building socialism" or "defending socialism." (See Figure 13.4.) The sub-
sequent U.S. actions designed to obstruct the Soviet actions Moscow
considered building or defending socialism were then connected to the act-
category of "pursuing world domination."
The U.S. Cold War script also has five acts. Atomic efforts were conceived
as a way to "show U.S. strength." This conceptualization began after the
Alamogordo test in 1945, when Truman was at the Potsdam conference.
U.S. economic and political efforts in Western Europe were the act of "build
a free society." Economic and military efforts, combined with restraint in
their aggressiveness toward the Soviet Union, were cast as "defending the
West." Soviet political, economic, and military efforts in Eastern Europe
were viewed as "totalitarian aggression." And elements of Soviet political
and military restraint were conceived as "caution." The U.S. script matrix
is presented in Figure 13.5.
A subject perceives an historical structure as a chain of recurring instances
of the same script. The perceived script defines the situation for the subject.
Over time, the script becomes "the way things are," reified as a natural
or traditional order. Yet historical structures endure only so long as they
are reiterated by the subjects involved.
Decision: Social Act Schemata

In the process of decision in this simulation, action flows not from choice
but from the subject's past and present construction of situations as instances
Figure 13.4: Soviet Act Construction in Rounds 1 & 2
Round 1:
build defend build world prudent
socialism socialism capitalism domination restraint
PPC PPC, ANC PUR PUR, BAB NAD
Round 2:
build
socialism
~.,,
support
I against
PPD
!-------
defend kJuard \under- against
socialism \utilize
PPD, ACD
-+---
build gain
capitalisn support
TE
pursue
world
!against !against r-
dominatiol1 I
BP, BAB,
TE I
I
I
prudent
restraint
NAD I
II ~~er-
tilize
----~~-----· ·-· -- __ _..1_~-----------
of scripts. The subject's psychocultural structure contains a variety of action

rules, encoded in the language of acts, which trigger themselves when
certain acts are perceived. Each repetition of such an action rule is dependent
on perceived satisfaction of its goals subsequent to its previous action.
There is no rational cognitive center making choices, rather, there is a
decentralized system of action rules whose conditions of survival make the
subject intelligent.
The Swiss cognitive developmental psychologist Jean Piaget proposes
that human action is guided by action schemata. 10 Schemata are dynamic
memory structures that identify stimuli in various situations, attach goals
to those stimuli, choose actions to achieve the goals, and-crucially-require
reinforcement by the success of their actions in attaining their goals.
Schemata record their most recent stimuli, goals, and actions, and then
monitor for goal achievement.
The Piagetian notion of action schema can be extended to the social
realm. A social act schema contains a stimulus consisting of a set of acts,
an act to be performed, and a goal consisting of acts. Social act schemata
are exogenously specified in this model; the stimulus, act, and goal of each
schema is designated at the start. What distinguishes a social act schema
320 Sanjoy Banerjee
Figure 13.5: U.S. Act Construction in Rounds 1 & 2
Round 1:
show build defend totalitarian caution
strength free West aggression
society
BAB PUR BAB, PPC ANC
PUR
Round 2:
show against
strength
BP
BAB
- - c---
build gain- differentiate,
free support against
society
TE
defend under- guard against

West utilize
BAB, TE
NAD
1-------
totalitarian against against
aggression
PPD
-
caution under-
ANC utilize
from other kinds of action schemata is that the former's stimulus and
reinforcement are produced by other schemata like itself.
The concept of social act schema allows a model of decision and action
that avoids the assumption of unbiased prior knowledge of choices and
their consequences by the agent. The question for a schema-guided agent
is not which single alternative to choose, but which subset of schemata
will enact themselves. Subjects must construct the stimulus in the current
situation and have constructed goal-achievement after the previous enactment
of the schema for the schema to be reenacted.
The U.S. act of "defend the West" is produced by a schema with the
two-act stimulus of Soviet "totalitarian aggression" and the U.S. act of
"build free society." The goal of the schema is Soviet "caution." The
meaning of this schema is that when the United States perceives the alarming
combination of its building free societies in Western Europe and Soviets
engaging in totalitarian aggression in the East and promoting subversion in
the West, they are motivated to enact the defense of the West in the belief
that such action will cause Soviet "caution."
A social act schema must be reinforced before it repeats its act. After
a schema is enacted, the subject monitors the next round for the goal. If
the goal is achieved, the schema is reenacted at the next instance of the
stimulus, otherwise not. A subject's schema will not repeat if any other
subject's schemata generate actions not conforming to the requirements of
the goal. So after enacting the "defend the West" schema described above,
the United States awaits evidence of Soviet caution. When it receives it,
the schema is reinforced and it repeats at the next stimulus. This process
was the fundamental mechanism of the early Cold War. Each superpower
interpreted the restraint of the other and caused by its own military buildup
and its own political consolidation of its zone of control. This reinforced
the remilitarization and bloc-building. In the Soviet case, bloc-building included
the suppression of noncommunist forces in Eastern Europe.
Once a social act schema generates an act to be enacted, an action
must be produced. An action is chosen whose intentional relations with
actions in the previous round conform to the requirements of the act. The
program takes the action from the appropriate intentional relations matrix
of the appropriate pair of rounds.
Social act schemata can be distinguished in discourses. The stimulus is
the conditions, described in the language of acts, that require the immediate
performance of the act. The goal is expressed as the desired acts by others
expected to result from the focal act. The reinforcement of a social act
schema arises from the causal attribution of goal satisfaction to the earlier
act under the circumstances of the stimulus.
The U.S. schema of "show strength" had the stimulus of Soviet totalitarian
aggression and the United States' building a free society in Western Europe,
and the goal of Soviet caution. The schema was present in the secret
discourse of the National Security Council. The actions of building atomic
bombs and proposing to maintain the U.S. monopoly through the Baruch
Plan were important constituents of the act of showing strength. NSC 30
(September 10, 1948) discusses atomic weapons in terms constitutive of
the social act schema in question. The "caution" goal of the schema is
expressed by:
[The Soviets] should never in fact be given the slightest reason to believe that
the U.S. would even consider not to use atomic weapons against them if
necessary. lt might take no more than a suggestion of such consideration,
perhaps magnified into a doubt, were it planted in the minds of responsible
Soviet officials, to provoke exactly the Soviet aggression which it is funda-
mentally U.S. policy to avert. 11
An implication of this passage is that the purpose of the U.S. atomic arsenal
is to avert Soviet aggression by threat of its use "if necessary."
The next sentence of NSC 30 expresses two aspects of the schema: the
stimulus and reinforcement from success in achieving the goal.
322 Sanjoy Banerjee
If Western Europe is to enjoy any feeling of security at the present time,

without which there can be no European economic recovery and little hope
for a future peaceful and stable world, it is in large degree because the atomic
bomb, under American trusteeship, offers the present major counterbalance
to the ever-present threat of Soviet military power. 12
This passage expresses the stimulus by referring to the background fact

of Western European recovery attendant to the U.S. act named here "build
a free society." The "ever-present threat of Soviet military power" invokes
the other part of the stimulus-Soviet "totalitarian aggression." The statement
"it is in large degree because (of] the atomic bomb" constitutes a causal
attribution of Soviet restraint until that date to the U.S. atomic arsenal-
to the United States' show of strength. As discussed above, such attribution
of goal-achievement to the act of a schema is the expression of its
reinforcement.
The Soviet schema for its act of "defending socialism" had the stimulus
of the U.S. quest for world domination, and the goal of inducing "prudent
restraint" by the United States. Elements of it were expressed in Zhdanov's
1947 Cominform speech.
Just as in the past the Munich policy untied the hands of the Nazi aggressors,
so today concessions to the new course of the United States and the imperialist
camp may encourage its inspirers to be even more insolent and aggressive.
The Communist Parties must therefore head the resistance to the plans of
imperialistic aggression along every line-state, economic and ideological. 13
Communists are exhorted to use every line, including state police power,
to achieve the goal of discouraging "imperialistic aggression" and its supposed
local agents. The stimulus of a Nazi-like U.S. quest for world domination
is also expressed. Further, if concessions "encourage" insolence and ag-
gressiveness, then the goal of the act of defending socialism must be to
discourage aggressiveness-to induce prudent restraint.
Reproduction of Subjects
Subjects are modeled here as psychocultural systems of perception and
decision. Subjects are systems of meaning that process meanings, thus
subjects are cultural subsystems. This is what allows groups, and not only
individual persons, to constitute subjects. The sharing of a subject-culture
by a group constitutes it as a subject. Scripts and schemata define subjects.
With scripts and their intentional relations matrices, reflective perception
arises. With social act schemata, reflective decisions are biased by the
subject's construction of the situation.
Different interacting subjects in a historical structure agree upon the

description of all actions and on the intentional relations matrices between
the actions of consecutive rounds. They do not share scripts or their acts.
Thus for the United States, its two actions of building atomic bombs and
not deploying them are joined into the act of defending the West nonag-
gressively. For the USSR, the same actions are separated into two different
acts. U.S. building of atomic bombs is combined with the Truman Doctrine
speech in the third round into the act of pursuing world domination, and
the action of not deploying the atomic bombs in Europe is assimilated to
the act of prudent restraint, which is seen in Moscow's social act schema
as a product of its own strength. Each subject has a unique clustering of
actions into acts.
Historical subjects are contingent, open to change under certain circum-
stances. The historical continuity of perception requires the reproduction
of scripted structures; the recognition of each act requires the earlier
perception of certain other acts. The survival of social act schemata requires
they be repeatedly stimulated and reinforced. If the subjects' scripts and
schemata are not in an appropriate relation to each other, the required
stimuli and reinforcements for the social act schemata of the initial structure
will not be produced, the scripts will not be recurrently perceived, and the
subjects will not be reproduced.
Conclusion: The Reproduction of Historical Structures

Subjects in a historical structure, even if they are in conflict, reproduce
together. Each is what it is because the others are what they are. Interacting
subjects in a historical structure are different and unequally powerful but
not separate, not innocent of each other.
Certain combinations of scripts and social act schemata yield reproducing
historical structures. For reproduction to occur, each social act schemata
must be stimulated and reinforced by the actions produced by the other
schemata. Such stimulation and reinforcement requires the actions to be
construed as acts within scripts. Hence reproduction requires the continuity
of reflective perception. The schemata must produce the right combinations
of actions so that the actions can have the required intentional relations to
the actions/acts of the previous round. Due to the reflective nature of
subjects, the model must be started up with the exogenous specification
of two initial rounds of action and with the actions of the first initial round
construed as acts. The process of reproduction is illustrated in simplified
form in Figure 13.6.
The reproduction model explains the temporary stability of historical
structures. The starting premise of the reproduction approach is that social
stability is not self-explanatory, but rather is the result of vulnerable cyclic
324 Sanjoy Banerjee
Figure 13.6: The Reproductive Process

(Perception) SUBJECT A (Decision)
/
Scripts stimulate, Schemata
reinforce
~. reflective
act-construction enact
~Actions/
reflective /
act-construction
~ enact
Scripts~ stimulate,
reinforce
~hemata
(Perception) SUBJECT B (Decision)
processes. The simulation shows how the right combination of scripts and
schemata can cause itself to be reiterated and yield a stable pattern of
social interaction. At the level of actions, there can be no more than a
chronology of events. At the level of acts, there can be stability and continuity.
The wrong combination of scripts and schemata fails to reproduce.
A distinguishing feature of this model is that while it is entirely rigorous,
it dovetails closely with some existing empirical methodologies. Structures
of meaning like acts and social act schemata are necessarily communicated
in historical discourses, and they can be discovered as objects in those
discourses. Patterns of action can be recovered from structural historiography.
An advantage of the present approach is that the descriptions of such
historiography can be preserved to a large degree in the formal model. 14
Notes
1. Robert Keohane. 1988. "International Institutions: Two Approaches." International
Studies Quarterly. December.
2. For further discussion of social reproduction, see Sanjoy Banerjee. 1984.
Dominant Classes and the State and Development: Theory and the Case of India.
Boulder, CO: Westview Press; Sanjoy Banerjee. 1986. "Reproduction of Social
Structures: An Artificial Intelligence Model." Journal of Conflict Resolution Volume
30, Number 2; Sanjoy Banerjee. Forthcoming. "Reproduction of Subjects in Historical
Structures: Emotion, Identity, and Attribution in the Early Cold War." International
Studies Quarterly.
3. Kenneth Waltz. 1979. Theory of International Politics. Reading, MA: Addison-
Wesley.
4. Christopher Achen and Duncan Snidal. 1989. "Rational Deterrence Theory and
Comparative Case Studies." World Politics Volume XLI, Number 2.
5. Robert Jervis. 1988. "Realism, Game Theory and Cooperation." World Politics
Vol. XL, Number 3.
6. Rom Harre. 1980. Social Being. Oxford: Basil Blackwell.
7. Waiter LaFeber. 1973. The Dynamics of World Power: A Documentary History
of United States Foreign Policy 1945-1973, Volume 11: Eastern Europe and the Soviet
ilnion. Arthur Schlesinger, general editor. New York: Chelsea House Publishers, p.
352.
8. Ibid., p. 355.
9. Arthur Vandenberg (ed.). 1952. The Private Papers of Senator Vandenberg.
Boston: Houghton Mifftin, p. 382.
10. Jean Piaget. 1971. Biology and Knowledge. Chicago: University of Chicago
Press. Jean Piaget. 1977. The Development of Thought: Equilibration of Cognitive
Structures. New York: Viking Press.
11. Thomas Etzold and John Lewis Gaddis (eds.). 1978. Containment: Documents
on American Policy and Strategy, 1945-1950. New York: Columbia University Press,
p. 341.
12. Ibid., p. 341.
13. LaFeber 1973, p. 360.
14. lt is not possible to reproduce the computer program that implemented the
model because of space considerations. The program was written in Prolog, and
the program and the rounds comprising the program's "runs" are available from
the author.
Bibliography
Achen, Christopher, and Snidal, Duncan. 1989. "Rational Deterrence Theory and
Comparative Case Studies." World Politics. Volume XLI, Number 2.
Banerjee, Sanjoy. 1984. Dominant Classes and the State and Development: Theory
and the Case of India. Boulder, CO: Westview Press.
- - - . 1986. "Reproduction of Social Structures: An Artificial Intelligence Model."
Journal of Conflict Resolution Volume 30, Number 2.
- - - . Forthcoming. "Reproduction of Subjects in Historical Structures: Emotion,
Identity, and Attribution in the Early Cold War." International Studies Quarterly.
Clocksin, William, and Mellish, C. 1984. Programming in Prolog. Berlin: Springer-
Verlag.
Etzold, Thomas, and Gaddis, John Lewis (eds.). 1978. Containment: Documents on
American Policy and Strategy, 1945-1950. New York: Columbia University Press.
Harre, Rom. 1980. Social Being. Oxford: Basil Blackwell.
Jervis, Robert. 1988. "Realism, Game Theory and Cooperation." World Politics Vol.
XL, Number 3.
326 Sanjoy Banerjee
Keohane, Robert. 1988. "International Institutions: Two Approaches." International

Studies Quarterly. December.
Lafeber, Waiter. 1973. The Dynamics of World Power: A Documentary History of
United States Foreign Policy 1945-1973, Volume 11, Eastern Europe and the Soviet
Union. Arthur Schlesinger, general editor. New York: Chelsea House Publishers.
Piaget, Jean. 1971. Biology and Knowledge. Chicago: University of Chicago Press.
___ . 1977. The Development of Thought: Equilibration of Cognitive Structures.
New York: Viking Press.
Waltz, Kenneth. 1979. Theory of International Politics. Reading, M.A: Addison-Wesley.
14
Theoretical Categories and Data Construction
in Computational Models of Foreign Policy
David J. Sylvan, Stephen J. Majeski,
and Jennifer L. Milliken
This chapter reports on an important phase in our project of constructing

a computational model of war recommendations. In other papers, we have
laid out the rationale of the project, along with some important features of
foreign policy recommendations that preclude a top-down, rule-based model.
More recently, we have moved toward the development of an alternative
computational model by specifying its conceptual and constitutive features
[15, 16, 23, 24, 25). We are now in the final stages of data construction;
when we finish, we will be able to begin writing and evaluating machine
code that embodies the computational model.
The specific issue addressed in this chapter is how to move from archival
materials to model-specific data, without in the process losing the critical
nuances by which policy recommendations are differentiated from each
other. Although we have some confidence in our approach and procedures,
they must be considered as tentative until further tested.
The Model
Our goal is to model the relations among the component parts of what
we call a bona fide policy recommendation. We use these terms ("component
parts," "bona fide") to stress that our modeling concern is constitutive:
what is necessary for a statement (written or oral) to be understood as a
genuine recommendation, rather than as something else. Accordingly, our
emphasis is cultural: how, within a particular foreign policy bureaucracy,
certain statements are fitted together into a comprehensible recommendation.
The work in this chapter was partially supported by NSF grant #SES-8520259.
327
328 D. J. Sylvan, S. J. Majeski, and J. L. Milliken
We deliberately eschew the cognitive components of this process, not because

they are uninteresting but (a) because our concern is with social, not
individual, activity; and (b) because for the most part, cognitive processes
are inaccessible to us. Many of the recommendations we are studying were
made by individuals who now are dead, and there simply is no way to
retrieve sufficient cognitive information to construct a plausibly valid model.
Before proceeding to a description of the model proper, it is worth noting
two features of such models in general. First, they must of necessity take
into account the construal within a particular culture of certain statements
as arguments, evidence, conclusions, and so forth. Categorizations of this
sort, however, are likely to be tacit rather than explicit; as we will argue
below, this leads us to adopt a strategy of "grounded theorizing." Second,
a focus on cultural definitions rather than cognitive processes means that
constitutive models of this type-understood as sets of complementary
relations among component parts-are not in most senses temporal (though
they may include memory). We do not claim to address the social processes
in which recommendations are put together by individuals or groups; nor
do we claim to look at how, over some period of time, a recommendation
comes to be understood as such. For purposes of presentation, we shall
lay out our model as a path; but this metaphor ought not to be taken
literally. Later in our project, we shall construct a genuinely processual
model that accounts for how certain recommendations flow from or are
circumscribed by others. At present, our concern is with how, for a given
bureaucratic and political culture, various statements are taxonomically related
to each other so as jointly to compose a bona fide policy recommendation.
The essence of the constitutive model is a path that connects descriptions
of a situation with particular policies. 1 Let us begin with situation descriptions.
These fall into three categories: global labels that pertain to the situation
as a whole; ceteris paribus consequences that set out what will happen if
policy is unchanged; and features of the situation that highlight particular
facets of what is going on. Situation descriptions serve as starting points
to the recommendation process; together with current proximate goals, they
can be understood as inputs on which computations are made. 2 Situation
descriptions (particularly situation features) are matched against current
proximate goals to see if the latter are being satisfied. That match permits
determination of the success (or lack therof) of the current policy. Situation
descriptions may also be matched against the likely outcome (itself a situation
description) of a policy being proposed by others. This second type of
match determines whether the current situation would be ameliorated by
the proposed policy; the criterion used is the current proximate goal. 3 Both
types of match are made by ascertaining the relevance of features to the
current or proposed policy.
Theoretical Categories and Data Construction 329
Unless the current or proposed policy is assessed as likely to work in

the future, the relevance of the features (usually reasons for failure or results
of success) is used to posit new proximate goals. These latter, in turn, are
matched against analogies, both for that country in the past (prior cases
in country), and for other countries (prior cases out of country), in which
similar new proximate goals were being pursued. The first analogy that
matches satisfactorily is then scanned for the policy that was employed.
Policies consist of tools and missions; the former are extracted from the
analogy and tailored so that their mission offers some chance of achieving
the new proximate goal. Sometimes analogies are used in assessing policies
proposed by someone else.
The model thus has six theoretical categories (situation descriptions,
relevance of features to current or proposed policy, proximate goals, tools,
missions, and analogies), three of which (situation descriptions, proximate
goals, and analogies) have subcategories. In principle, then, a recommendation
could comprise up to ten components. The data construction problem is
to extract from archival and other sources as many of those components
as possible for each recommendation. The theoretical categorization problem
is to "ground" the categories and their interrelations as firmly as possible
in the participants' understandings of their statements. Each problem pre-
supposes the other; they cannot be disentangled.
First Steps: Entering Documents

and Abducing Categories
Our model was originally developed as a result of intensively scrutinizing
the documents pertaining to U.S. policy toward Laos in 1961. In its current
form, the model reflects a similarly intensive scrutiny of documents pertaining
to U.S. policy toward Vietnam in 1964-1965. The purpose of the model is
to be able to account for the content of each of the principal recommendations
made with respect to both of these countries in both time periods. 4 We
estimate that there were around ninety such recommendations in total. In
order to come up with this number, to develop the model, and to arrive at
a point where data construction and theoretical categorization could begin,
we had to take several preliminary steps.
First, archival materials had to be collected. In general, they fell into
three categories: documents from the Kennedy and Johnson presidential
libraries; documents quoted or referred to in the Pentagon Papers and other,
smaller compilations; and speeches, news reports, reminiscences, and other
materials from the Congressional Record, newspaper articles, memoirs,
histories, and interviews. All told, we amassed some 3,250 items, ranging
anywhere from a one-sentence news report to a 200-page military report. 5
As we collected the documents, we entered them into a machine-readable

document flow. This involved pulling out (usually in verbatim form) the
various statements that seemed to embody aspects of recommendations or
outlooks6 in a document and typing them into a text file. 7 Each document
was entered in chronological order and was given both standard identifying
information (e.g., author, addressee, date) and a series of codes. 8 Appendix
1 shows the entry for one (rather famous) such document; note particularly
its codes. All told, the document flow takes up around 8 megabytes of disk
storage space. As might be imagined, the document entry was time-
consuming: lt took two entire years of twenty-hour weeks to complete.
Our next step was to use the codes to pull out entries pertaining to a
particular topic such as troop recommendations and put them in a separate
file. We scanned longer entries and extracted only those portions relevant
to the topic. Sometimes we found it necessary to use more than one coding
category; at other times we found that some of the entries sharing a certain
code did not appear relevant to a particular topic, or that we needed to
search for entries marked by a combination of codes. The resulting "sifted"
lists were considerably shorter than the document flow but still too large
to use as an everyday tool. 9 Accordingly, we summarized those lists into
yet another file, this one small enough so that a printout could easily be
scanned and flipped through by hand. 10
The point of producing such summaries was to be able to page back
and forth through all of the entries for a given decision. In so doing, we
began to develop a feel for a decision as a whole. 11 After studying the
summaries for a period of time, we were able to discern patterns in the
data: which documents seemed to go together, who was allied with whom,
and how the agenda seemed to change over time. At this point, we were
able for the first time to identify tentatively the major recommendations
that were put forward during the course of that decisional period. With this
initial specification of whole recommendations, we were in a position to look
closely at them, both in summarized and (from time to time) in document
flow form. We began to hunt for patterns within recommendations: how
recommendations were put together and what some of the standard forms
were that they displayed. In this way, we began to sketch out the features
of a model; to posit other, complementary features; and to go back to the
recommendations to see if those features were present. Eventually, we were
able to produce the original version of our constitutive model. When we
repeated the process for a second decisional period, we were able to discern
where modifications of the original model were needed so that it could be
generalized across both periods. At this point, we were in a position to
begin the actual tasks of data construction and theoretical categorization.
The process we have described thus far has several aspects worth noting.
First, the documents and other archival sources from which we worked
reflect a variety of different perspectives, depending on whom they were

written by and when they were written. The participants arrived at a set
of interpretive positions by which they construed documents as types of
speech acts. Had this not occurred, there could have been no dialogue, no
agreements, and certainly no conflict between positions understood as
opposing ones. One of our tasks preparatory to data construction has been
to grasp this set of interpretive positions-not necessarily to agree with
them (in fact, the opposite has usually occurred) but to ground our modeling
efforts.
Second, it was only after identifying these positions that we were able
to arrive at a preliminary identification of cases of recommendations. Which
statements seemed to be understood as a recommendation? Where did one
recommendation end and another one start? Were two different statements
understood as embodying essentially the same recommendation or as
expressing complementary recommendations or as evidence of sharply
opposing recommendations? Answering these questions in a preliminary
fashion permitted us to form a first, rough set of cases. 12 This set we
further refined (see below); it was only then that we began to select among
cases. At this point, and throughout much of the data construction process,
the bigger problem was to identify cases than to select among them. For
us, cases do not come prepackaged or given in some simple fashion by a
set of theoretical concerns. Rather, they are identified by the analyst on
the basis of participants' conceptual categories; for both observer and
participant, the effort is more akin to construction than to sampling.
Third, once we had identified recommendations as such, we were in a
position to identify various of their features (e.g., global labels) whose contents
could be assessed as similar or different. Such features are, of course, the
theoretical categories we have been referring to; we postulate their existence
as a way of accounting both for the relations (similarity, difference; but also
temporal ordering) between cases and for the construed coherence among
the various statements that make up a case. We can further refine the
theoretical categories in order to account for differences between groups of
recommendations across (putative) decisional periods. To the extent that
the postulated categories help solve these various theoretical problems, we
are well on our way to building a constitutive model of how different
statements are in fact bona fide cases. To put the matter in epistemological
terminology, we have abduced theoretical categories [19).
In general, our methodology bears a close resemblance to what Glaser
and Straus [7) describe as the generation of "grounded theory." Our concern
is to "generate conceptual categories or their properties from evidence"; to
abduce theoretical categories that account for the differences between cases;
to specify those categories as precisely as possible; and to generate relations
between categories. By contrast, we have tried to avoid a priori specifications
of a model or of relevant data, not because we lacked advance theoretical

ideas about how foreign policy recommendations are made or because we
had no prior notions about the O.S. role in Vietnam, but because we wanted
to make sure that the model really did take into account both the nuances
and the principal features of the documents.
Of course, there is always a danger that one can drown in information.
At times, we certainly felt like that. But by continually asking theoretical
questions, especially about how the bureaucratic participants packaged
information in the process of argumentation, we were able to cope with the
immense number of documents. We found that in our successive passes
through the documents that we could begin to formulate categories if we
took copious notes on the "story" that we would tell about the documents
at that moment, similar to the techniques of qualitative fieldwork. At times
we felt like anthropologists studying a fairly primitive tribe. 13
Next Steps: Constructing and Categorizing Data

The basic issue in data construction is to specify the theoretically relevant
components of each recommendation. Similarly, the basic issue in theoretical
categorization is to ground the categories and their interrelations in the data.
These processes required several steps.
First, we needed to correct the summaries constructed from the document
flow. At times, a number of theoretically important features had not been
given sufficient emphasis in a given summary, which required adding elements
(e.g., detailed situation features) to the entries in the summary. On occasion,
however, we could see holes in our initial specification of the recommendations;
we had omitted important documents or events that, while not coded for
the decision, were nonetheless germane. Accordingly, we revised the sum-
maries, adding a considerable amount of additional information. 14 In certain
instances, the process of revision led us to reassess the accuracy of our
summarization, and so we also corrected some of our wording.
We then took the revised summaries, studied them again, and divided
them into phases. Specifically, we demarcated certain time periods of the
summary as being taken up with a debate over a particular main policy
line. The phase opened when the policy line was articulated for the first
time as the principal element on the agenda; the phase closed when a
reasonably final decision was made about the policy line. During a phase,
there were also subsidiary policy lines put forward, mostly in reaction to
the main line. By definition, main lines were contiguous: There was one at
each moment of a decisional period. However, while there were usually
several subsidiary lines at any moment, they did not necessarily run the
entire length of the phase; they could also overlap with each other within
a phase.
Phasing-as a particular kind of categorization-was an activity that

required considerable discussion among us. Since we did not always agree
on the identity of the principal issue on the policy agenda or on which
documents counted as part of which line, we needed to put forward plausible
accounts to each other of the bureaucratic arguments that had occurred at
a particular time. In so doing, we began to acquire a deeper understanding
of what the participants in a particular debate meant when they used certain
phraseology and at whom they were aiming their comments. As we convinced
each other that participants shared specific understandings, we were further
able to ground our theoretical categories in the conceptual categories of the
various bureaucrats and politicians we were studying.
What we found, at least for the troops decisions, is that at any given
time there were three policy lines: a main one and two subsidiary ones.
Other policy positions were of course put forward, and on rare occasions
they were discussed at high levels; but as a general rule, the president and
his top advisers simply did not consider more than three principal lines at
a time. On the other hand, there were rarely instances in which a phase
was marked by only one or two principal recommendations. We can think
of various reasons for this "rule of three," ranging from cognitive limitations
to the functionality of interorganizational distinctions. However, it is important
to point out that although the three lines could be arrayed along some sort
of "hardness" dimension, the standard ways in which "hardness" is normally
treated in the foreign policy literature [1, 5, 9, 10, 11, 12, 21, 22] do not
seem to have been borne out in our data construction efforts. It is not the
case that the main line is usually the middle-of-the-road position or that
particular individuals or organizations are usually identified with particularly
hard or soft positions or that recommendations of some particular degree
of hardness or softness usually come to the fore (much less win out).
In general, we find that the standard antinomies in the foreign policy
literature-hard and soft, idealistic and pragmatic, closed and open-are
not very applicable to our documents. At best, these categories need
considerable nuancing if they are to capture the operative distinctions drawn
by the participants; at worst, the standard categories are serious hindrances
to our data construction efforts. The problem here is not that the foreign
policy literature is insufficiently detailed, since some of the classic studies
in the field are finely grained. Nor is the problem necessarily that we are
studying recommendations while most of the literature is on decisions; it
should, after all, be possible to analyze decisions by means of subtle
distinctions without retracing the steps by which those decisions are framed.
Rather, the relative inapplicability of standard categories to our work appears
to stem from our adoption of a grounded research strategy instead of a
routine hypothesis-testing one. 15
The phasing completed, we could turn to the question of identifying

particular documents (or speeches in meetings) as articulating a particular
line of a particular phase. This was not difficult to do, although it was quite
tedious, especially as we tried to identify (and later use) every document
or speech associated with a given line. Numerous documents that seemed
at first blush to be advocating a particular line turned out, on closer
inspection, to be discussing alternatives or advocating a different or minor
line. Often, there were coded messages present: An argument against one
position turned out to be understood by the participants as an argument
in favor of another position. Occasionally, we would find that a given
bureaucrat changed sides and shifted from one line to another during a
given phase; at times this occurred subtly, via omitting certain catchphrases
that had until then been a routine part of a particular line.
The next step was to take each document 16 identified as belonging to a
particular phase-line combination and to pull out the portions of it that
pertained to the different categories posited in our model. Ideally, there
would be ten such categories per document, all of them springing to our
eyes. In practice, we found that certain categories were difficult or impossible
to glean from a particular document. The reasons for this varied. Sometimes,
a category might not be germane to a recommendation and we would be
unable to infer its contents or otherwise posit its existence. If a given adviser
(or, on occasion, the president) thought that a particular policy would soon
bear fruit, there would be no reason for him to come up with analogies. 17
Alternatively, a category might be tacit: unspecified because of the exigencies
of form (short cables) or the demands of comity or sycophancy (e.g., not
to criticize the president). Such tacit categories are prevalent in many
recommendations; it took considerable piecing-together on our part to tease
them out of the documents.
When we extracted categories from recommendations, we arrayed them
by phase-line combinations. An example is shown in Appendix 2. The phase
is the second in the troop decision, characterized by a main policy line18
of sending troops to the central highlands. Appendix 2 shows one of the
subsidiary policy lines during that phase: bomb the North harder rather
than sending troops to the South. Note that square brackets are our
interpolations of tacit categories in the documents.
The next step, which we are now in the process of undertaking, is to
amalgamate the entries under each category in each phase-line combination.
Amalgamating entries from different documents should force us to eliminate
redundancy and leave us with one item per entry; if it does not, it suggests
that the policy line will have to be further divided into sublines to preserve
the nuances that differentiate recommendations. These sublines occur from
time to time when advocates of a particular position change their emphasis
over the course of a phase. Sublines also occur on occasion when a subsidiary
line contains both negative injunctions and positive statements. For example,
the line shown in Appendix 2 appears to contain two recommendations: to
step up the bombing and not to send troops. Though the recommendations
are closely linked, they are discrete. 19
Appendix 3 shows an example of an amalgamated phase-line combination.
Specifically, the appendix shows the amalgamated entries for the main line
(bombing) of the first phase of the troops recommendations. 20 Note that
the entries are for the most part worded as indirect quotations from the
documents.
Our aim in following this strategy was to preserve the participants' most
common linguistic categories while at the same time providing us with the
flexibility to eliminate terms we considered redundant or tangential. We
consider recommendations to be abstract phenomena that can be concretized,
or instantiated, in various ways. The specific enunciations contained in the
various documents associated with a given line serve as so many concre-
tizations of that line. By constructing our amalgamated entries out of indirect
quotations, pruned of redundancies and tangents, we can put together a
more direct, essential concretization of a specific line. In this way, the
contents of the different theoretical categories in a recommendation stand
out more sharply than they do in many of the actual documents.
The final stage in data construction will come when we take the
amalgamated entries and rewrite them in a syntactical form amenable to
list manipulation. Our model emphasizes the role of verbal pointers-for
example, the reasons for the failure to achieve a current proximate goal
point to what the new proximate goal should be-and so we will want to
write the entries in a way that will facilitate the extraction, transformation,
and gluing-together of those pointers. lt is important to emphasize, however,
that the rewrites will only alter the syntax of the entries, and then only for
programming purposes; the semantics will stay constant. As we have
emphasized elsewhere [25], we will be the ones determining the success or
failure of matches between verbal pointers, not some standardized English-
language or bureaucratic lexicon. Our intention is to build a computational
model that preserves shades of meaning, not obliterates them.
Further Analyses
The primary result of the data construction and theoretical categorization
will be a set of objects used to generate a computational model of policy
recommendations. As that model is completed, we shall turn to a related
(and even more difficult [26)) task: building a dynamic model that accounts
for the ways in which recommendations accompany or follow on each other.
These tasks ought not-at least at some point-to preclude us from analyzing
the data objects in their own right. There are two sorts of such "spin-off"
analyses that come to mind.
One is a kind of bureaucratic sociology. We can take the various phase-
line combinations in their full (not amalgamated) form and look to see if
there are recurring alliances between individuals or agencies. The literature
on U.S. foreign policy abounds with cultural, economic, organizational, and
political arguments in favor of such alliances [3, 9, 14, 20], and the data
objects we will have at our disposal provide one means to assess such
arguments. At first glance, the data and the "rule of three" seem to support
the existence of such alliances, though we will need considerably more
phase-line combinations before we can be sure about this.
The second sort of analysis we can undertake is to put forward an
internal classification scheme by which certain types of recommendations
(not policy lines, but groups of policy lines) can be distinguished. A very
preliminary look suggests, for example, that recommendations by the military
tend not to appeal to analogies, while recommendations to continue the
status quo tend not to specify ceteris paribus consequences. Regardless of
whether these particular findings hold up, any systematic connection between
the content of recommendations and their form opens the possibility that
there exists a well-understood grammar of recommendations. If that turns
out to be the case, then we would have a second avenue (in addition to
models such as we propose above) to understanding the limits of foreign
policy: what kinds of recommendations can and cannot be made.
The generation of a grounded computational model and the possible
accomplishment of these tasks offer attractive rewards for the imminent
completion of the final steps of data construction. They are, as it were,
the light at the end of the tunnel.
Appendix 1:
An Entry in the Document Flow
RUSK-VN
RECOMMENDATIONS:
1. US OBJECTIVE: The central objective of the US in SVN must be to ensure
that NVN not succeed in taking over or determining the future of SVN by force.
2. We must accomplish this objective without a general war if possible.
3. [follows outlook 12) MAIN COURSES OF ACTION: a) Maximum SVNese effort.
We must use the leverage of US presence and assistance to insist that the SVNese
leaders declare a moratorium on their bickering and knuckle down to the increased
effort needed to defeat the VC. They must be told bluntly that they can't take us
for granted but must earn our help by their performance.
b) A level of US effort in SVN, as a supplement to the best the SVNese can do,
to deny a VC success.
c) Maintenance of present character of air actions against NVN, both for its
practical effects and to establish that there is no sanctuary for participants in the
war.
d) The elimination of the sanctuary is the principal brake upon direct Chinese
participation. We shouldn't, for the present, attack targets in the immediate Hanoi-
Haiphong area. Priority should be given to any need for air strikes on targets in
SVN.
e) Intensify the mobilization of "other flags" in support of SVN.
f) Pursue our readiness to seek a peaceful solution through any available channel.
g) Initiate as soon as possible the "Acheson Plan" in SVN; if not for the entire
country, perhaps in the Ill and IV corps, or in selected provinces around Saigon.
OUTLOOK: 1. US OBJECTIVE: a) The central objective of the US in SVN must
be to insure that NVN not succeed in taking over or determining the future of SVN
by force.
b) We must accomplish this objective without a general war if possible.
c) The "war aim" of the US is not concerned with what the SVNese would do
if they were left alone. There are many problems in the country that only the
SVNese can solve.
d) US forces are present in SVN only because of the aggression of Hanoi in
sending men and arms into the South. If this aggression were removed, US forces
could be withdrawn.
e) We wouldn't use US forces to settle issues in SVN among the Buddhists,
Catholics, the sects, the local Chinese and Cambodian communities, the Montagnards,
and the genuinely "indigenous" VC.
f) The sole basis for employing US forces is the aggression from the North.
2. US COMMITMENTS: There can be no serious debate about the fact that we
have commitment to assist the SVNese to resist aggression.
3. If the SVNese were to ask us to withdraw our help, we would have to do so.
There is no present likelihood that they will do so.
4. The integrity of the US commitment is the principal pillar of peace throughout
the world. If that commitment becomes unreliable, the comm world would draw
conclusions that would lead to our ruin and almost certainly to a catastrophic war.
5. So long as the SVNese are prepared to fight for themselves, we cannot abandon
them without disaster to peace and to our interests throughout the world.
6. COMMENT ON THE PRESENT SITUATION: There is no question but that
the situation in SVN is critical. It is said that we are "losing"; this means that we
aren't making headway, but rather falling behind, in the effort to stop the infil and
pacify the country.
7. But that doesn't mean that the VC are "winning"; they have the power to
disrupt, but they aren't capable of occupying and organizing the country or any
major part of it.
8. The VC can be denied a victory, even if complete pacif will be a long and
tortuous prospect.
9. THE RISKS: There are obvious risks in any engagement between free and
comm countries, especially where large comm countries are contiguous to the area
of conflict. But these risks are present for the comms as well.
10. If they discover that we are less resolved than they, the prospect for the
future is exceedingly dark. Moscow and Peiping don't wish a general war with us
over SEA. Our problem, therefore, is to deny to Hanoi success in SVN without
taking action on our side that would force the other side to move to higher levels
of conflict.
11. If they decide to move to a larger war rather than fail to absorb SVN, we
couldn't shrink from that eventuality; but such a decision on their part doesn't
appear likely.
12. It is least likely in relation to what we do in SVN.
13. [follows rec. 3b) Even present levels of US forces aren't yet reflected in
corresponding damage to the VC. Reinforcements now in course should open the
way to a war plan to engage concentrations of VC with punishing effect.
AUTHOR: Rusk
TO: Johnson
DATE: July 1, 1965
SUBJECT: US obj, GVN/US, troops, am resp, US out, Buddhist, GVN pot, peace,
geo pot, ally react, US stakes, critical, SVN disint, infll, pacif, VC, risk, esc, NVN
bomb, SVN bomb, flags, poli sol, inter text, prestige, metaphor, SEA
Appendix 2:
An Example of Categorized Recommendations
**B. Bombing, not troops, line
***1. Situation descriptions
••••a. Global labels
The chances of a turnaround in SVN remain less than even. McG Bundy,
McNamara, and Rusk, 3/6.
The outlook is indeed serious. Taylor, 3/11.
Downward trend. Taylor, 3/18.
"**b. Ceteris paribus consequences
Unless NVNese support is checked, SVN mil and paramil resources increased,
pacif goals and concepts refined, admin effectiveness improved, and an adequate
pot-psych base created, little hope of stemming the tide of VC insurgency. Taylor,
3/11.
••••c. Features of the situation
Losing ground at increasing rate in cntryside in Jan and Feb; morale lifted; no
ev yet new GVN can turn things around; Hanoi not yet persuaded Hanoi to leave
its neighbors alone. McG Bundy, McNamara, and Rusk, 3/6.
Basic unresolved problem is provision of security for pop; due to lack satisfactory
progress destroying VC in cntryside, to continuing capab of VC to replace losses
and incr in strength (due in part to infll, which can't be stopped on ground), and
to our inability to establish and maintain an effective govt. Taylor, 3/7.
Steady growth in VC mil capabs; continuing erosion of GVN position in cntryside;
and general lack significant progress in pacif. VC have cut country in two. Extensive
infll. War-weariness and lack of confidence in ultimate defeat of VC. Taylor, 3/11.
Improved morale of SVNese; they have glimpse of light at end of long and
tortuous tunnel. Taylor, 3/12.
Morale of ARVN in north coast very low; troop discipline there poor; VC control
most of province. Taylor, 3/13.
GVN trained mil manpower shortage. Deteriorating situation in I and 11 Corps
areas. Taylor, 3/18.
Moment of relative govt stability and some renewed spirit. Rusk, 3/23.
More stable GVN, better SVN mil morale. McCone, 3126.
NVN attitude has hardened. McCone, 412.
***2. Relevance of features to current or proposed policy air strikes have lifted
morale and made beginning toward persuading Hanoi to leave its neighbors alone.
Have intl and US opinion on our side. Pentagon has been concentrating on mil results
against guerrillas in field, when should have been concentrating on police control
from village up. McG Bundy, McNamara, and Rusk, 316.
[Since many of problems caused by support of insurg from Hanoi, bombing
remains both critical and the most promising tool; success at it could make pacif
efforts more successful.] Taylor, 3/7.
Only US resources can provide the pressures on the N necessary to check Hanoi's
support. Internal measures and programs required will require marked increase in
US support and partic. Taylor, 3111.
Air strikes on N and in S, together w I landing of Marines, have improved morale
SVNese and given them glimpse of light at end of long and tortuous tunnel. Taylor,
3/12.
We're not losing because of insufficient mil forces but because of poor performance.
Taylor, 3/13.
We will soon have to decide whether to get by with inadequate indigenous forces
or to supplement them with third-country troops, largely if not exclusively US. But
introducing US division in highlands would place our forces in area with highly
exposed LOCs to coast, serious logistical problems, Montagnard separatism accu-
sations. Putting troops in enclaves is inglorious static defense. More generally, putting
US troops in SVN increases involvement, exposes greater forces, invites greater
losses, raises sensitive command questions w I GVN allies, may encourage them to
slack off, appear as colonizer, and considerable doubt number of ARVN relieved
would have great significance in reducing the manpower gap. Taylor, 3/18.
Moment of relative govt stability and some renewed spirit resulting from strikes
against NVN. Rusk, 3123.
Air strikes have been moderately successful from a mil point of view. McNamara,
3126.
Air strikes have stabilized GVN and improved SVNese mil morale. Sustained
strikes may prompt Hanoi to offer negots. McCone, 3126.
Even if we make little progress or go backward in pacif during coming months,
delay will be inconsequential if [because of bombing], Hanoi throws in sponge and
agrees to cut off VC. Would need massive ground forces to have any short-term
effect. Taylor, 3127.
Quat not persuaded of necessity US trps. Anti-Americanism just under surface.
Three divs high; limited absorptive capacity of SVN; logistical limitations. Taylor,
3/29.
Strikes have if anything led to hardened NVN attitudes because too limited in
targets. [Ground troops obviously can't do anything to affect NVNese attitudes since
don't cause them any direct pain; instead, they'll get mired in the jungle.) McCone,
4/2.
•**3. Proximate goals
**••a. Current proximate goals
This is a Vietnamese war in which we are helping in areas where the VNese
cannot help themselves. Taylor, 3/2.
We must continue to make every effort in the pacification area. McG Bundy,
Make maximum effort in 1965 to raise new forces and improve use of those
already in being. Continue trying to get Hanoi to stop infll. Taylor, 3/7.
[Diminish Hanoi's will to continue support of VC.) Taylor, 3/11.
(Is a VNese war); don't get them into the "let the US do it" attitude. Taylor, 3/
18.
Make a determined effort to increase our effectiveness and that of the GVN.
Rusk, 3/23.
[Get Hanoi to offer negots.) McCone, 3/26.
Get Hanoi to throw in the sponge. Taylor, 3/27.
Force Hanoi to call off VC. McCone, 4/2.
••••b. New proximate goals
[Shift pacif effort to police control at village level.) McG Bundy, McNamara, and
Rusk, 3/6.
Engender official and popular confidence VC can and will be defeated. Taylor, 3/
11.
Improve morale of SVNese. Taylor, 3/12.
[Improve morale of ARVN.) Taylor, 3/12.
•**4. Tools
[Continue bombing; push hard on more effective pacif tools.) McG Bundy,
[Keep up bombing and keep plugging at pacif.) Taylor, 3/7. Continue and increase
bombing of North; create pol and psych programs to bring populace behind GVN's
programs. Taylor, 3/11. [Continue bombing.) Taylor, 3/12.
Don't send troops. Taylor, 3/18.
Except possibly for some aspects of the coordination process in the pacification
program, we don't need to revise main structure of our nonmil effort. Rusk, 3/23.
[Continue sustained bombing.) McCone, 3/26.
[Continue bombing. Don't send troops.) Taylor, 3/26.
[Don't send troops.) Taylor, 3/26.
(Bomb much harder and faster; don't send troops.) McCone, 4/2.
•••5. Missions
Convince Hanoi price it must pay is too costly to bear. [Bolster ARVN morale
and cut down on VC recruiting.) Taylor, 3/11.
Win the loyalty and support of the people by convincing them that the GVN is
interested in and working for their security and welfare. Rusk, 3/23.
[Hurt Hanoi enough that it offer negots.) McCone, 3/26. [Hurt Hanoi in all kinds
of targets.) McCone, 4/2.
***6. Analogies
••••a. Prior cases in Vietnam
[None)
••••b. Prior cases other than Vietnam
Negative analogy: old French role of alien colonizer and conqueror. Taylor, 3/18.
Appendix 3:
An Amalgamated Phase-Line Combination
Phase 1: Bombing Line
1. Situation descriptions
a. Global labels
Situation in SVN deteriorating and substantially likely to come completely apart
in next two months.
b. CPCs
Without new US action, defeat appears inevitable.
c. Features of the situation
Unstable GVN. Infiltration problem. VC gaining in countryside.
2. Relevance of FS's to current or proposed policy
Current policy hasn't helped get stable GVN and thus hasn't been able to solve
problems of infiltration and VC gains. Current policy also is limited and thus can't
scare NVN into calling off insurgency.
3. Proximate goals
a. CPGs
Stable GVN. Harass NVN.
b. NPGs
Raise actual and anticipated cost to NVN and demonstrate US will, so that NVN
call off VC.
4. Tools
Graduated bombing of NVN.
5. Mission
Slow, gradual escalation: firepower, targets, locale, so that NVN will become
increasingly scared.
6. Analogies
a. Inside SVN
Negative analogies: Tonkin Gulf, because episodic. Bien Hoa and Brinks, because
not respond at all.
Positive analogies: Previous bombing plans from spring 1964: bombing induce
NVNese to call off insurgency in SVN.
b. Outside SVN
Negative analogies: French in Vietnam, US in Korea (at end): inconclusive and
messy.
Positive analogies: Korea 1950, Italy 1948, Berlin: firm.
Notes
1. This is a highly condensed summary; the model is discussed at length elsewhere
[16]. The current version of the model differs from the original version in minor
respects (e.g., standard operating procedures are replaced by in-country and out-
of-country analogies), each such difference prompted by a Jack of success in our
attempts at constructing data for some particular category of the model. The
summary in this chapter omits such vital details of the model as branching and
extraction mechanisms.
2. In this discussion, we pass over the long-term goals for policy relative to a
particular country. Elsewhere [25], we have argued that those long-term goals can
be understood as dictated by a country's place characteristics.
3. This second type of match is usually done in the process of constructing
"negative recommendations"-not to do something. This is a feature we did not
have in the original version of the model, largely because the Laotian case we used
in constructing the model had only one negative recommendation.
4. Specifically, we will have five sets of recommendations: those that eventuated
in the acceptance of a Geneva-brokered coalition government in Laos; those that
eventuated in the policy of sending large numbers of advisers to South Vietnam;
those that eventuated in a bombing campaign over North Vietnam; those that
eventuated in the policy of sending large numbers of combat troops to South
Vietnam; and, as a control, those that eventuated in the policy (in June 1950) of
large-scale ground fighting against North Korea.
5. lt is worth noting that an enormous number of documents remain classified.
Still others are unclassified but unknown to us. Many recommendations were made
orally, so that there would be no records. Thus, there is no way that we can ever
have a complete list of documents on Vietnam and Laos in 1961 and 1963-1965.
(We begin the second period with the summer of 1963, since that was when plans
for the coup against Diem-which opened the way to more extensive military
activities against the North-were begun.) The best we can say is that, at the
moment, we seem to have most of the germane documents in the public domain.
6. When documents were first entered in the flow, we sorted statements (usually
sentences or groups of sentences) into two categories: "recommendations" and
"outlooks." These classifications were useful for scanning long entries in the document
flow. However, our goal is to see how individual statements are linked together into
bona fide recommendations; toward that end, we reserve the term "recommendation"
for these bona fide wholes. By contrast, when we wish to refer to the crude
bookkeeping categorization of individual statements as recommendations rather than
outlooks, we will use the term "recommendation statements."
7. At the time the document entry was carried out, there were no low-priced
optical scanners that could substitute for typing the documents. Even if there had
been such scanners, and even if they could have read faint, repeatedly photocopied
copies of cables and handwritten notes, much of the document entry involved
rearranging the text (to highlight recommendations and outlooks), summarizing it,
and inferring the existence and nature of other documents or events referred to in
the text. We doubt that any time savings would have resulted from scanning the
documents.
8. The codes were devised by Milliken, who also constructed the document flow.
A codebook is available from the authors. lt should be noted that the codes vary
widely: They range from apparent topics covered in a document (e.g., troops, bombing
of North Vietnam, pacification) to various metaphors in or rhetorical features of a
document (e.g., game playing, racial terminology, sexual imagery).
9. The sifted list for those "troops" entries that plausibly pertained to the July
1965 decision to send combat forces to Vietnam was about 800 kilobytes in length;
it ran from August 1964 to July 1965. lt should be noted in passing that assembling
this list was highly computation-intensive, involving a fast microprocessor, large,
random-access memory, and a big, high-resolution display monitor to facilitate
scanning and cutting and pasting long documents.
10. The original summary for the troop decision was about 200 kilobytes in
length; when printed out in rotated format, two columns to a page, in eight-point
type size, it took up some thirty pages.
11. For that matter, the process of producing the summaries led to a first level
of familiarization as a result of looking at the same documents time and again.
12. Actually, there were two sets for each decisional period: one for the Laos
1961 recommendations and one for the Vietnam ground war recommendations in
1964-1965.
13. In fact, it had not been our intention originally to adopt this methodology.
We had thought that we would simply be able to take the documents and fit them
into some preset categories, the latter clumped as families of rules. lt took us two
years of pushing and pulling the documents so that they would fit those categories,
along with another year of generating an ever-proliferating number of subsidiary and
higher-level rules, before we began to employ a "grounded" strategy. As might be
expected, we could not carry out even this latter strategy without occasionally
running into blind alleys. We have not included blow-by-blow accounts of all our
decisions-not so much because we wish to spare ourselves embarrassment as
because our goal in this chapter is to lay out a reconstructive logic for the methodology
we eventually adopted.
14. In the troop decision summary, the length of the file increased by 25 percent.
15. In this regard, it is interesting to note that the basic chronologies we
constructed for various decisional periods are much closer to the accounts in good
bureaucratic histories of the Vietnam War [6,8, 13,28) than to the accounts in more
avowedly social scientific studies of the same events [10, 11,27).
16. At times, we were able to use the summaries instead of the actual documents;
normally, though, we needed to look at the detailed wording in the documents
themselves. On occasion, we found it useful to work from photocopies of the original
documents, as they provided us with a better sense of paragraph structure than we
could obtain in the document flow.
17. Analogies are usually produced "to order," as the needs of an argument
demand. This does not mean they are not important in the construction of
recommendations, but rather that they do not come prepackaged with a whole series
of features that can be plucked like arrows from a quiver. (This "prepackaged"
view is widespread in the foreign policy literature and is shared by scholars whose
views are otherwise quite disparate [2,4, 12, 17, 18,29).) Instead, we ftnd that analogical
features are adduced in the process of argumentation, regardless of how they may
first be thought up. We should note in passing that high-level bureaucrats rarely
produce more than one or two analogies; there appears to be a bias against
"cleverness."
18. Eventually rejected by the president, though revived successfully several
phases later.
19. In support of this position, it is interesting to note that when the situation
worsened (the US. embassy being attacked) and it looked as if bombing would not
do anything in the short run for the South, the advocates of the bombing and not
troops line switched to supporting coastal enclaves as a limited, controllable measure-
unlike their evaluation of sending troops to the highlands.
20. Were we constructing data for the group of recommendations that eventuated
in the policy of systematically bombing North Vietnam, the entries shown in Appendix
3 would need to be disaggregated into separate lines and phases. Advocates of
bombing, for example, differed sharply over the goals, the extent, and the timing
of the bombing. If, however, our interest lies in constructing data for the recom-
mendations that eventuated in the policy of sending large numbers of combat troops
to South Vietnam, the distinctions among different bombing recommendations can
be disregarded. When serious discussions first began to take place about whether
and how to send troops to Vietnam, what was relevant to the participants about
such issues is that two troops recommendations (sending them to establish a cordon
sanitaire and sending them to protect U.S. bases) were put forward in partial
opposition to various schemes for bombing. When participants focused on troops,
disputes over bombing policies were more or less ignored.
Bibliography
(1] Graham T. Allison and Morton H. Halperin. Bureaucratic politics: A paradigm
and some policy implications. In Raymond Tanter and Richard H. Ullman, editors,
Theory and Policy in International Relations, pages 40-79. Princeton University
Press, 1972.
[2] Paul A. Anderson. Justification and precedents as constraints in foreign policy
decision-making. American Journal of Political Science 25(4):738-761, 1981.
[3] Richard J. Barnet. Roots of War. Atheneum, 1972.
[4] James P. Bennett. Data stories: Learning about learning from the U.S. experience
in Vietnam. In Donald A. Sylvan and Steve Chan, editors, Foreign Policy Decision
Making: Perception, Cognition, and Artificial Intelligence, pages 227-279. Praeger,
1984.
[5] Lloyd S. Etheredge. A World of Men: The Private Sources of American Foreign
Policy. MIT Press, 1978.
(6] William Conrad Gibbons. The U.S. Government and the Vietnam War: Executive
and Legislative Roles and Relationships. Two volumes. Princeton University Press,
1986.
[7] Barney G. Glaser and Anselm L. Straus. The Discovery of Grounded Theory:
Strategies for Qualitative Research. Aldine, 1967.
(8] David Halberstam. The Best and the Brightest. Random House, 1972.
(9) Morton H. Halperin. Bureaucratic Politics and Foreign Policy. Brookings Institution,
1974.
(10) Ole R. Holsti and James N. Rosenau. American Leadership in World Affairs:
Vietnam and the Breakdown of Consensus. Alien & (Jnwin, 1984.
[11] lrving L. Janis. Victims of Groupthink. Houghton Mifftin, 1972.
[12] Robert Jervis. Perception and Misperception in International Politics. Princeton
(Jniversity Press, 1976.
[13) George MeT. Kahin. Intervention: How America Became Involved in Vietnam.
Alfred A. Knopf, 1986.
[14] Gabriel Kolko. The Roots of American Foreign Policy: An Analysis of Power
and Purpose. Beacon Press, 1969.
[15) Stephen J. Majeski and David J. Sylvan. Decision and recommendation heuristics
for war involvement: The 1961 Vietnam commitments. A progress report. Paper
presented at the annual convention of the International Studies Association,
Washington, D.C., 1985.
(16) Stephen J. Majeski and David J. Sylvan. A constitutive model of the a.s. foreign
policy recommendation process. Paper presented at the annual meeting of the
American Political Science Association, Chicago, 1987.
[17) Dwain Mefford. Analogical reasoning and the definition of the situation: Back
to Snyder for concepts and forward to artificial intelligence for method. In Charles
F. Hermann, Charles W. Kegley, Jr., and James N. Rosenau, editors, New Directions
in the Study of Foreign Policy. pages 221-244. Alien & (Jnwin, 1987.
[18) Richard E. Neustadt and Ernest R. May. Thinking in Time: The Uses of History
for Decision-Makers. Free Press, 1986.
[19) Charles Sanders Peirce. Philosophical Writings, pages 150-156. Edited by Justus
Bucher. Dover, 1955.
[20) Franz Schurmann. The Logic of World Power: An Inquiry into the Origins,
Currents, and Contradictions of World Politics. Pantheon Books, 1974.
[21] Glenn H. Snyder and Paul Diesing. Conflict Among Nations: Bargaining, Decision
Making, and System Structure in International Crises. Princeton (Jniversity Press,
1977.
[22) John G. Stoessinger. Crusaders and Pragmatists: Movers of Modern American
Foreign Policy. W. W. Norton, second edition, 1985.
[23) David Sylvan and Stephen Majeski. A formalization of decision and measurement
heuristics for war initiation: The 1961 Vietnam commitments. Paper presented
at the annual meeting of the American Political Science Association, Chicago,
1983.
[24] David Sylvan and Stephen Majeski. Problems of contextual specificity in
computational models of foreign policy. Paper presented at the annual convention
of the International Studies Association, Anaheim, 1986.
[25] David Sylvan and Stephen Majeski. Intervention and neutralization: The 1961
decisions in Vietnam and Laos. Paper presented at the annual convention of the
International Studies Association, Washington, D.C., 1987.
[26) David J. Sylvan, Stephen J. Majeski, and Jennifer L. Milliken. Modeling policy
spaces: Some preliminary issues. Paper presented at the annual convention of
the International Studies Association, London, 1989.
[27) Wallace J. Thies. When Governments Collide: Coercion and Diplomacy in the
Vietnam Conflict, 1964-1968. University of California Press, 1980.
[28) United States, Senate Committee on Foreign Relations. Committee Print: The
U.S. government and the Vietnam war: Executive and legislative roles and
relationships: Part Ill: January-July 1965, by William Conrad Gibbons. tOOth
Congress, second session, 1988.
[29) Vaacov V. I. Vertzberger. Foreign policy decisionmakers as practical-intuitive
historians: Applied history and its shortcomings. International Studies Quarterly
30(2):223-247, 1986.
15
Semantic Content Analysis: A New Methodology
for the RELATUS Natural Language Environment
John C. Mal/ery
Classification is the fundamental interpretive process found in political

cognition, whether individual or collective. Recall that Ronald Reagan's
"freedom fighters" and Leonid Brezhnev's "bandits and criminals" were two
classifications for the same Afghan resistance, designed to call forth desired
political actions in their audiences. Examples such as this suggest that
classification is a fundamental political act (Mefford, 1988a) standing behind
conflict and cooperation.
When a problem is classified, or "framed," free from temporal anchors
in the present, this classificational process shades into analogical reasoning,
or precedent logic, which has been proposed to model political understanding
and organizational decision making (Aiker and Christensen, 1972; Alker,
Bennett, and Mefford, 1980; Mallery and Hurwitz, 1987; Mefford, 1987;
Mallery, 1988a: 38-47). Because classification is a central cognitive process
with high relevance for politics, tools to simulate classification are a high
research priority. They can help reveal the cognitive dimensions of the
phenomena as they yield methods for computational politics-the subfield
of political science that uses symbolic reasoning techniques from artificial
intelligence to formally model politics.
This chapter introduces semantic content analysis, a methodology whose
vehicle is automatic recognition and classification of instances in the knowl-
edge representations of texts, or text models. 1 Here, the RELATUS Natural
Language System (introduced in Alker et al., this volume) provides an
environment that allows an analyst to syntactically parse and semantically
represent a variety of English texts, allowing for the identification of
classifications being made. A semantic content analysis of texts results.
Earlier computerized content-analytic techniques2 could suggest the thematic
orientations expressed in particular textual corpora, but they could neither
reliably identify the sign (positive or negative mention) nor determine the
347
348 John C. Mallery
direction of value or goal orientations. Instead of using word frequencies to

indicate thematic orientation, semantic content analysis examines explicit
structural relationships and inherent interpretations in text models.
Semantic content analysis differs from traditional content analysis because
it operates on referentially integrated text models. Referential integration
means that references to the same object or relation, which may appear
in different sentences of a text, are resolved and represented as the same
semantic node. In successive steps, text models are created incrementally
through a process that syntactically analyzes each sentence and constructs
a logical form. The reference system then uses the logical form to merge
intersentential coreferences into referentially integrated semantic represen-
tations. After syntactic analysis captures structural relations within sentences,
referential integration compiles the information across sentences. Both are
critically important for political texts because they make explicit who does
what to whom throughout a text.
There are three phases to semantic content analysis:
• Text Representation: The sentences of a text are syntactically parsed

and semantically represented to create referentially integrated text models
• Classification: The political analyst applies lexical recognizers, 3 designed
in advance, to locate and classify relational configurations in the text
model
• Inspection: The analyst uses any number of interfaces for inspecting
the text model to view the classifications
Relevance of Text Modeling

Computational politics has two branches: applications that build models
or knowledge representations from artificial languages, such as rule-based
expert systems, and those that begin from natural language. Since most
human knowledge is encoded in text, most models will draw from textual
sources. If the interpretative process of data making is to be explicit and
itself an object of study and debate-rather than hidden in forgotten data-
coding practices-political analysts must use natural language AI systems
to create knowledge representations from the source texts (Duffy and Mallery,
1986; Mallery, 1988a: 47-66; Duffy, 1988, 1989; Alker et al., this volume).
Drawing knowledge directly from long-term social memory (text) into
computer systems (Lenat and Guha, 1990), provides one of the few ways
to increase the productivity and capability of social science. Beyond these
epistemological considerations, two general propositions about ontology of
political phenomena argue that computational politics requires a text modeling
capability:
Semantic Content Analysis 349
• Symbolic Basis of Political Rationality: Individual cognition is grounded

in "procedural rationality" (Simon, 1985); a symbolically mediated,
functional account of human reasoning processes. The purposeful
cognitive processes of individuals and organizations, dependent on the
linguistic interpretations of meaning and intentions, account for choice
and provide the source of political action (Lasswell, 1927, 1935; Lasswell,
Leites et al., 1949; Deutsch, 1953; George, 1959; Pool, Lasswell, Lerner
et al., 1970; Alker, 1975, 1984, 1986; Habermas, 1979, 1981; Dallmayr,
1984; Shapiro, 1984, 1988; Taylor, 1985; Mallery, Hurwitz, and Duffy,
1987; Mallery, 1988a; Mefford, 1988b; Duffy, 1987, 1989; Alker et al.,
this volume; Hurwitz, 1990).
• Social Systems as Patterns of Conversations: The embeddings of
people within discourse communities is the main constraint and stimulant
of political cognition (Wittgenstein, 1953; Deutsch, 1953; Habermas,
1979, 1981; Dallmayr, 1984; Harre, 1985; Alker, 1986; Mallery, 1988a).
Within this framework, political communication and scientific discussions
about politics become patterns of arguments. Where consensus is
possible, there are shared norms of valid argument. In conflict situations,
the validity of argumentative moves and even facts become problematic.
Reading the Chapter

The chapter is organized into two major divisions. The first section
provides a summary overview of the theory and practice that converts texts
into referentially integrated knowledge representations. The second section
reviews the implemented lexical classification system, explaining how to
define lexical recognizers, organize them into hierarchies, and apply them
to text models. lt also reviews the classifications systems developed so far.
A discussion of applications draws on the theoretical underpinning of
RELATUS and analogizes from traditional content analysis to develop some
notions of reliability and validity for semantic content analysis. lt also
proposes other analytical applications, such as precedent logics, modeling
differential perceptions, and recognizing argument connectives. The final
application domain explains how lexical classification can extend referential
coverage in text models, yielding better representations for analysis.
Text Representation in RELATUS

The RELATUS System4 acquires knowledge by mapping "literal and
explicit" texts into a dynamic knowledge representation that captures their
referential structure. This representation lends itself to the regeneration of
surface texts and to inferential operations associated with commonsense
reasoning (Minsky, 1987), such as analogy and generalization. Some major
features of RELATUS are:
350 John C. Mal/ery
• Text .Modeling Environment: It provides a powerful and general foundation

for text modeling, including facilities for research and simulation of
cognitive processes, whether individual, collective, or abstract. RELATUS
was originally conceived as a "social science workbench" to model
politically relevant texts and support actual or counterfactual simulations
(Duffy and Mallery, 1986; Mallery, 1988a; Alker, Duffy, Hurwitz, and
Mallery, this volume).
• Domain Independence: It is a domain-independent AI system because
its bottom-up operation relies on the ontology (form) of knowledge
representation instead of its specific content (Mallery, 1988a).
• Phenomenological Representation: It initially represents meanings by
taking snapshots of grammatical relations from sentence parses and
referentially linking the constituents across sentences. The initial picture-
like, or eidetic, representation can receive subsequent, and possibly
divergent, interpretations. Thus, difficult interpretation problems, such
as the analysis of belief operators, are deferred. 5
• Hermeneutic Orientation: It is grounded in a constructivist theory of
meaning that composes interpretations from an initial eidetic repre-
sentation rather than a decompositional theory that reduces sentences
to putatively universal (but cognitively implausible) semantic primitives.6
• Conservation of Information: It does not lose any information present
in the original texts; as interpretation proceeds, information is only
added. For example, grammatical relations are carried forward from
syntactic analysis into the semantic representation.
• large Texts: It is capable of processing large texts (to date, up to
several hundred pages) quickly (about one minute to parse and reference
250 sentence texts, or roughly ten pages, on Symbolics XL 1200 LISP
machines). This means that once texts have been converted to "literal
and explicit" English, around 450 to 600 pages per hour could be
parsed and referenced.
A clear understanding of the implemented processing model is crucial

for:
• creation and analysis of text models with RELATUS

• determination of the validity of semantic content analyses
• recognition of the contemporary limitations of the technology
• research to extend hermeneutically oriented text modeling technology
Lexical-Interpretive Semantics
Semantic perception is the process of mapping from a syntactic rep-
resentation into a semantic representation (Mallery and Duffy, 1990). Tra-
ditionally, universalist semantics (Katz and Fodor, 1963; Schank, 1972; Schank
and Abelson, 1977) advocates determining equivalent meanings (paraphrases)
through the decomposition of different surface forms to a canonical semantic
form composed of semantic universals, such as "conceptual dependency"
primitives (Schank and Abelson, 1977). But, lexicalist semantic theories
argue that most meaning equivalences must be determined constructively
for specific linguistic communities (or even individual language users) and
dynamically for the intentional context. Experimental psycholinguistics sup-
ports the lexicalist position (Fodor et al., 1980; Gentner and Landers, 1985).
As a central feature of strategic language and decision, referential opacity
poses a debilitating dilemma for semantic universalism. Opaque contexts
are linguistic situations where statements are scoped by belief-suspending
constructions, such as potentially counterfactual verbs of belief, intention,
or request. Verb tense or aspect indicating future occurrences, subjunctive
mood, or conditionals have the same effect as do adjectives like "imaginary."
Opaque contexts require an understander to independently determine the
referential status of their contents. Semantic universalism's perceptual ap-
paratus, discrimination nets, provides no means of identifying opaque contexts
in order to avoid merging equalities (or semantically canonizing possibly but
not necessarily equivalent paraphrases) across them without prior deliberation
(Maida and Shapiro, 1982; Mallery, 1987). Addition of this capability would
require a representation of surface semantics before decompositional per-
ception-but that obviates the need for a universalist representation!
In RELATUS the construction of semantic representations from canonical
grammatical relations and the original lexical items (word stems) is informed
by a theory of lexical-interpretive semantics. Lexical-interpretive semantics
assumes that meaning equivalences arise because alternative lexical real-
izations accomplish sufficiently similar speaker goals to allow substitution.
A practical argument for dynamically determining meaning congruences is
the intractability of a static analysis with sufficient details and nuances to
capture subtle variations in speaker goals.
Instead of relying on static equivalences determined in advance, lexical-
interpretive semantics requires identification of meaning equivalences at
reference time. Although the theory calls for such dynamic determination
of meaning equivalences at reference time for historical, individual speakers,
the present practice in RELATUS relies on a universal syntax for idealized
language users and, to the extent implemented, static meaning congruences
for belief systems with specific background knowledge.
Since RELATUS retains the original words from sentences, the resulting
semantic representations are lexicalist, and referential opacity is the norm
(Maida and Shapiro, 1982). Substitution of equals across opaque contexts
never takes place; instead, identity relations may be asserted after valid
equalities are determined. Although the determination of dynamic meaning
352 John C. Ma/lery
congruences may require attention to opaque contexts, at least, the opaque

contexts and their contents are represented so that a reasoner may consider
them!
Lexical-interpretive semantics is hermeneutic because the theory em-
phasizes interpretation based on the effective-histories of language users
and the intentional structures of communicative situations. 7 The RELATUS
system reaches the level of eidetic representation and even somewhat beyond.
Lexical, constraint-based graph matching forms the ground for more complex
deliberative processes of a more general and open-ended hermeneutic in-
terpretation. In theory, lexical categories should be induced for each un-
derstander's history (Piaget, 1979).
Models of Natural Language Processing

Along a referential bootstrap sequence, natural language processing models
range from single-sense, literal and explicit processing through multisense,
literal and explicit processing to partially deliberative, and ultimately, fully
deliberative or Al-complete processing. 8
• Immediate Reference simply looks up literal relations from explicit

representations in memory.
• Deliberative Reference relies on reasoning, which may be arbitrarily
complex, to select possible referents.
Single sense refers to word usage where only one meaning of the word
is used within any of its parts of speech (e.g., noun, verb, adjective). Multi-
sense processing requires the ability to discriminate different senses of words
within part of speech and recognize equivalent paraphrases. Litera/language
refers to text in which no metaphors or other tropes appear and all words
are used according to a single definitional authority, such as an analyst or
a text producer. Explicit language contains no implicit premises or referents
that require inferences drawing on background knowledge. Because the
referential connections between sentences are explicit, the coherence of the
text becomes manifest.
Immediate reference accounts for the most basic model, single-sense,
literal and explicit processing. Constraint-directed graph matching finds the
correspondences from sentences to semantic memory. Word senses cause
no confusion because only one sense appears for each part of speech. The
absence of tropes and implicit referents defers some of the difficult problems
of reference.
The move to multiple word senses begins by using lexical classification
to recognize and label different senses. lt continues by using meaning
congruences to recognize dynamically equivalent paraphrases for references
across sentences. The present version of RELATUS is now approaching

these capabilities. The reference system has the ability to recognize nom-
inalizations of verbs. The lexical classifier uses selection constraints to identify
different senses of words. Identifying equivalent paraphrases and differentiating
word senses shades into deliberative reference as the application of selection
constraints and cross-graph mappings demand inferences. Although this
model is far from unrestricted natural language, it is nevertheless a useful
beginning that allows creation of interesting and useful representations for
text.
The important scientific point is that the distinction between immediate
and deliberative reference provides a theoretical framework to decompose
Al-complete natural language understanding into a bootstrap succession of
simpler, independent extensions that lead cumulatively to better processing
models. There is no claim that a system "understands" natural language;
rather, grammatical, referenetial, and inferential competence gradually evolves
from highly restricted toward ever less restricted language as incremental
research discovers how to extend the implemented model.
Creating Text Models

RELATUS gains broad coverage and domain-independence from a bottom-
up strategy that combines a general syntactic analysis with a constraint-
posting reference system to create large, referentially integrated semantic
representations. 9 The major components that make this possible will now
be reviewed.
Knowledge Representation. Unlike conventional knowledge representation
schemes, the GNOSCERE subsystem provides belief system objects, which
are each instantiations of a knowledge-base object-class, handling a set of
generic operations. 10 All processing in RELATUS is organized around belief
systems. Their major operations include syntactic analysis using the current
parser, semantic reference using the constraint-interpreting reference system,
sentential constraint posting, and question answering. Other operations include
tracking the dependencies of semantic structures to their source sentences,
displaying knowledge structures, processing user-defined directives (actions
for the system to take when given imperative sentences by the user), and
invoking predefined "if-added" procedures whenever trigger relations are
created in a belief system. Their semantic representations are a special
class of graph structure (or semantic network) constructed from tenary
relations or labeled binary relations. An important feature of tenary relations
is that they are arbitrarily expressive. 11 Different interpretive regimes can
be concurrently modeled by instantiating multiple belief systems and directing
them to parse and represent different texts. Within belief systems, memory
is implemented as a "society" of communicating graph-node agents (Minsky,
354 John C. Mallery
1987), which support over 500 operations. Belief systems maintain both
short-term and long-term societies of graph-node objects and manage the
locality of these knowledge structures to improve performance for large-
scale applications. 12 The process of text modeling is mapping surface text
into belief-system knowledge representations.
Constraint-Interpreting Reference. A constraint-interpreting reference sys-
tem (Mallery, 1990) is an interpreter and an extensible set of constraints.
The reference system finds and creates graph structure in semantic rep-
resentations, serving a function analogous to LISP's intern. 13 The constraints
constitute a declarative language for describing graph structures. Collections
of constraints are bundled in units called reference specifications (or ref-
specs).
Based on the lexical markers and bidirectional binary relations, the self-
indexation of GNOSCERE belief systems evolves as new tokens are added
and is extended by lexical classifications. Unlike most data base systems
that rely on fixed or partial indices, the dynamic extension of this full
indexication allows the reference system to be complete (find all the
possibilities) and correct (select the right ones).
This reference system finds graph nodes satisfying any ordinary constraint
specification, which include no inferences, in time independent of data base
size. To match a ref-spec, the system:
• Orders constraints according to their "pruning power" -an often precise

factor reflecting the number of nodes satisfying the constraint
• Generates a possibility space from the constraint satisfied by the fewest
nodes
• Applies constraints, previously ordered, to efficiently prune the possibility
space, yielding nodes satisfying the specification 14
This system is termed a reference system because it aims to computationally

model internal reference. It also allows users to define specialized constraints
for their applications. The constraint language presently contains about 150
types of constraints, a number of which are lexical predicates for use in
sentential reference. A meaning congruence facility currently finds referents
by constructing alternate reference specifications for certain regular cases
of agentive nouns (e.g., murderer) and simple derived nominals (e.g., de-
struction). In general, all other systems that access the belief-system
knowledge bases must express their data base operations as ref-specs.
Syntactic Analysis. The Duffy parser 15 (Duffy, 1987) is a deep-structure
transformational parser (Chomsky, 1965; Jackendoff, 1972) for English
syntax. 16 The parser produces a directed cyclic graph that describes the
syntactic constituents of a sentence. Interwoven within this cyclic graph
are structure-sharing trees (directed acyclic graphs) that describe the syntactic
deep structure and syntactic surface structure of a sentence, representations

of which constitute the parser's output. The procedural implementation of
the parser makes it very fast and therefore suitable for parsing large texts.
A key innovation that makes transformations tractable involves maintaining
clausal constituents in a list so that movement merely involves copying and
reordering the list rather than exponential, tree-descent copying of parse
structure. 17 The parser has quite broad coverage of English grammar.
The parser incorporates a lexical disambiguator (Duffy, 1986) that identifies
the correct part of speech for each word in a sentence by means of a
constraint propagation scheme (Waltz, 1975). When an ambiguous lexical
category is encountered, the disambiguator exploits constraints propagated
from its neighbors. Thus, the successful disambiguation of one ambiguity
in a sentence propagates additional constraint that may be helpful in
disambiguating other ambiguities. Unlike some parsers that handle multiple
parts of speech, the disambiguator and the syntactic analyzer are functionally
independent. This allows the disambiguator to run before the parser attempts
to analyze the grammatical structure of each sentence, making parsing more
efficient by removing from consideration alternate parses arising from lexical
ambiguity.
Sentential Constraint Posting. A sentential constraint poster converts the
output of the parser (parse graphs) into reference specifications, which
mediate the mapping from syntax to semantics. The construction of sentential
reference specifications draws grammatical relations from deep structure
and quantifies noun phrases based on surface structure (Mallery, 1985). This
process is called sentential constraint posting because the various pieces
of information composing the sentential description are incrementally collected
in a single sentential reference specification. When complete, the reference
specification is resolved for a belief system. As successive sentences of a
text are referenced, structures referring to existing representations are found
and new structures are added for those without referential antecedents. In
this way, the system builds up a referentially integrated semantic represen-
tation, a belief system graph structure that correctly captures references
to the same entities across sentences. This process of intersentential reference
makes possible the construction of referentially coherent text models on a
large scale.
Syntax Interface. The only language- or (linguistic) theory-dependent
components of RELATUS are the syntactic components, including the
preparser, the parser, the lexicon, the sentential constraint poster, and the
syntactic part of the sentence generator. Parsing sentences syntactically is
really decoding syntax into ref-specs, whereas generating sentences is
encoding ref-specs into syntax. 18 The recent installation of a new Common
LISP version of the Duffy parser provided the opportunity to encapsulate
the entire interface to syntax in a single interface object supporting a
356 John C. Mallery
declarative protocol. Thus, any parser-generator pair that can answer to the
protocol can provide the syntax service for RELATUS. Naturally, the parser-
generator pair needs to support a parser capable of tractably producing
sentential reference specifications 19 and, preferably, a generator, driven by
the same grammar, to invert structures produced by the parser. The syntax
interface spans all of the syntax-related commands throughout RELATUS.
The Text-Processing Cycle

This section illustrates the creation of referentially integrated text models
over three levels of linguistic processing:
• Syntax: Sentence strings are syntactically analyzed to produce surface

structure and necessary deep structure.
• Logical Form: The lexical content and structural relations of syntax
are expressed as sentential reference specification in the constraint
language.
• Semantics: The reference system creates referentially integrated graph
representation as resolves the constraint description, looking for matching
antecedents but creating new structures when none are found.
The syntactic analysis process begins when a belief system parses a

sentence stream from an editor buffer or text file. Before proceeding, the
belief system determines the current deictic context 20 either from global
values or from a more specific set of parameters associated with the sentence
stream. lt invokes a preparser that accepts a sentence from the stream,
produces a preparsed sentence list, and feeds it along with the deictic
context to the parser. Preparsing expands contractions, evaluates LISP forms,
and removes comments in the text stream. The parser returns an analyzed
sentence object and the belief system references the sentence. This process
first invokes the sentential constraint poster to convert the parse to a ref-
spec. Next, the ref-spec references itself with respect to the belief system.
The belief system repeats the procedure sequentially and incrementally until
it reaches the end of a sentence stream-a stream to any text source,
including files, buffers-regions, or the user.
Figure 15.1 shows a syntactic parse of a sentence from a RELATUS
parsable text about the 1956 Soviet intervention in Hungary. The dashed
arrow connecting "nominal-6" to "nominal-8" indicates the intrasentential
coreference found when the parser resolves the possessive pronoun "its"
to "USSR." This equality is then carried over through constraint posting.
Figure 15.2 shows the isolated semantic structure to which the reference
system resolves the sentence. The top-level relation reads: "government-3,"
"request-2," "instruct-1." Similarly, "instruct-1" reads "ON-Security-Council-
Figure 15.1: The syntactic parse of a sentence from a RELATUS parsable text about
the 1956 Soviet intervention in Hungary that was derived from the Butterworth (1976)
narrative. (Kern-2 is shown as closed.)
r--------------------------.
SENTENCE-600
IMre-Nagy's governMent requested that the
UN-Security-Council instruct the USSR to
negotiate its differences with
IMre-Nagy's governMent.
I," "instruct-I," "negotiate-3." Notice that the object of "negotiate-3" is

"difference-I," and it has a subject relation that is read "difference-I" "of-
79" "USSR-I." But the subject of "negotiate-3," "USSR-I" has an object
relation that is read backwards "difference-I," "of-79," "USSR-I" for human
intelligibility. In each case, "of-79" is the same relation, but it is accessed
from different directions. When it is an object relation, it appears in italic
font to distinguish it from its display as a subject relation in normal font.
This illustrates the bidirectional interpretation of GNOSCERE representations.
Figure I5.3 shows the ref-spec that mapped between the syntactic parse
in Figure 15.1 and the semantic structure in Figure 15.2. The gennames
(word + number) scattered around the ref-spec hierarchy correspond to
the gennames in the semantic representation of Figure 15.3. These corre-
spondences show the resolution of the constraints in the ref-spec. The top
(left) ellipse is the ref-spec for the main or matrix relation of the sentence.
The gennamed word inside the ellipse, "request-2," denotes the graph node
to which the ref-spec is resolved in the belief system. The ref-spec has a
subject ref-spec, resolved to "Government-2" and an object ref-spec, resolved
358 John C. Mallery
Figure 15.2: The isolated semantic structure to which the reference system resolves
the syntactic parse from Figure 15.1.
Subject
UN-SECURITY-CDUHCIL-1
( HRS-TENSE -299 ) - - - - ( PRESENT )
---1( H::::~::E-298
( HRS-RSPECT -291 )---'::::;::::::::=:.__
Object ) - ( PRESENT)
HECDTIRTE -3 ( HRS-RSPECT -289 )-(PERFECT)
Object
DIFFERENCE-I
HRS-CRRDINRLITY 18 -~
( COVERNnEHT -2 )
YD~~-----(USSR-1)
to "instruct-1." Between those ellipses is one labeled constraints. Below are

some constraints on request, such as the mandatory constraint TRUE. Below
TRUE is a PMSUBJECT-RELATION with the arguments "has-tense," "past."
This constraint says to prefer (P) a node for a "request" that has a subject
relation, "has-tense," whose object is "past." If the successful referent does
not satisfy this preference, the constraint mandatorily (M) assures it by
creating a subject relation for the referent. "Government-2," the object ref-
spec for "request-2," has a constraint INDIVIOOAL-P that requires the
referent to be an individual rather than a universally quantified node.
Existing Text Models

Text models in several content domains from international politics have
been developed using the RELATUS technology.
Butterworth Conflict Narratives. The first political text modeling at MIT
were the author's attempts to model the Butterworth (1976) summary account
of the 1956 Soviet intervention in Hungary using the Winston-Katz analogy
system between 1980 and 1983.21 By 1984 the author was able to parse
and fully represent in RELATUS both the Hungary story and the 1968 Soviet
intervention in Czechoslovakia. The texts each run about ten pages and
create about 6,000 graph structure nodes each. In 1987 an undergraduate
student parsed and represented several more Butterworth summaries, in-
cluding the 1954 U.S. intervention in Guatemala, the Bay of Pigs, and the
Jordanian-Lebanese civil war. In 1989 Hurwitz parsed and represented the
Figure 15.3: The sentential ref-spec, or logical form, that mediates reference from the syntactic parse of Figure 15.1 into the semantic structure
of Figure 15.2.
IMDIVIDUAL-P
Cons~r•ints
Re/-Sp•c
REGUEST-2 Conatr•i,.,ts
Canstr•l nt-LI st
SubJ•ct
UM-&ECURITY-CDUNCIL -I
ObJ•ct Coutr•ints
INSTIUCT-1
INDIVIDUAL-P
ObJet:t
MECOTIRTE-3
INDIVIDUAL-P
,_ _ • Constr«.ints
&UIJECT-REUTIDN Constr•ints
Constr•ints Con.traints
Const.r•lnt-List
Constraints PftSUIJECT -RELATION

Const~lnl-List H0-222
360 John C. Mallery
Butterworth account of the Cuban Missile Crisis for comparison to other

coding schemes (Aiker, Hurwitz, Mallery, and Sherman, 1989).
Prisoner's Dilemma Protocols. Hurwitz (1990) has parsed and represented
the largest amount of text (over 100 pages) yet processed by RELATUS.
He has entered protocols of sequential prisoner's dilemma (SPD) games,
originally acquired from a series of experiments conducted by Alker with
MIT undergraduates (Alker and Hurwitz, 1980). These protocols contain
narrative accounts of game play augmented by natural language statements
by players about their expectations, beliefs, and normative responses con-
cerning the course of play. His analysis shows that interactional structure
in the sequential prisoner's dilemma follows from self-interpretive and reflective
notions of appropriate conduct (Aiker and Hurwitz, 1980; Alker and Tanaka,
1981, Alker, 1985) rather than, for example, unreflective tit-for-tat reciprocity
(Axelrod, 1984).
The SHERFACS International Conflict Dataset. A feature vector editor
(Mallery, 1988b) can automatically generate RELATUS text models for any
cases selected from 700 international conflicts or 1,200 domestic disputes
since 1945 in the SHERFACS International Conflict Management Dataset
(Sherman, 1987a, 1987b, 1988). In April 1988 the author created a 250,000
node representation of fourteen SHERFACS cases. Later in August, a lexical
classification system was developed to identify Sherman conflict actions
categories (Mallery, 1988c).
Newspaper Articles on Vietnam. Devereux (1989) recently parsed and
represented a number of articles from major U.S. newspapers covering the
September 3, 1967, presidential elections in South Vietnam. He aims to use
semantic content analysis to test hypotheses of an "investment" theory of
the news media.
Additional Facilities
Semantic Inversion. Since semantic representations become very com-
plicated quickly, manual or ad hoc strategies for constructing reference
specifications from graph structures are fragile and error-prone at best. The
semantic inverter is an interpreter that can invert the function performed
by constraint-interpreting reference. In essence, semantic inversion allows
one to point at some graph structure and have a reference specification
constructed that will find structures like it. Given a seed node in the semantic
graph structure and some constraints to delineate a graph region for
incorporation, the semantic inverter traverses the graph structure by following
all relations from the initial node that satisfy the set of incorporation
constraints. Another type of inversion activity generates graphical displays
of RELATUS knowledge structures in the Belief Examiner Window while a
generation activity is used by an experimental sentence generator to create
surface sentences from semantic structure. 22
Question Answering. The development of the first semantic inverter in

1984 was motivated by the goal for belief systems to answer various types
of literal and explicit questions. The key insight conceives of the graph
structure created by referencing questions as a set of constraints connected
to some unknown (e.g., person, thing, place, truth-value, or enabling cause).
Semantic inversion of the representation for a question, beginning from the
unknown, constructs a reference specification that finds the answer plus
the question's unknown.
The present question-answering facility answers yes-no questions (including
tag questions) and wh-questions (except when). 23 Questions can be asked
about anything explicitly represented in a belief system-even things that
were never stated but were combined by intersentential reference or were
inferred by lexical classification. Figure 15.4 shows some examples of
RELATUS answering questions. In general, questions are correctly and
robustly answered to the extent that the question parses and references
correctly so that semantic inversion can build correct constraints that pick
out the right answer. For this reason, question answering is an effective
way to test the correct operation of the entire cycle that encodes text
models.
Editor Mode for Text Modeling. As an extension to ZMACS, the native
Emacs text editor of the LISP machine, the RELATUS editor mode features
over eighty different commands useful for preparing, parsing, and representing
text as well as verifying the semantic structures produced in belief systems.
There are enough commands in the RELATUS mode that new users need
not type any LISP forms to process their text. They can add information
to the lexicon, make their own private lexicon patch files, parse text, and
represent it in belief systems. They can change file and buffer attributes
for various aspects of the deictic context, such as the source, audience,
time, and belief system. They can use various mouse-sensitive inspectors
to examine the surface structure and deep structure of parses. They can
view the reference specifications produced by their sentences. They can
reference sentences in their chosen belief systems and examine the contents
of those belief systems. These kinds of facilities are important not just to
speed the task of representing text but also to make these tools available
to non-programmers or anyone who is not familiar with the implementation.
Belief-System Examiner. The belief-system examiner provides a graphical
display interface to the semantic structures of belief systems. It uses the
semantic inverter to walk regions of semantic structure and generate displays.
The examiner supports numerous display commands, including ones to show
structure created by a sentence (Figure 15.2), generalization hierarchies,
category instances ordered by creation time, and causal-intention structure.
For instance, all figures displaying semantic structure are created automatically
by a command that scales and hardcopies any displays. It also supports
Figure 15.4: Answering questions about the 1956 Soviet intervention in Hungary. Questions answering is a way to inspect the 1..1
text model and to see how lexical classification has extended the represented knowledge. ~
MORRISOI'I Q
Question to ask ARPA: (411D to end)
> Why did 1ft~e-nagy fell f~o" political powe~?
I'IORRISOI'I Q
FALL-~'• causes ere: PUT-1
SENTENCE-267
FALL-7
Question to ask ARPA: (GII> to end)
> ~ho was eKecuted?
I'IORRISON Q
For the question, EKECUTE-5, I find: II'IRE-I'IAGV-1 ASSOCIATE-3
SENTENCE-268
EKECUTE-5
Question to ask ARPA: (411) to end)
> Did the UI'I-General-Asseftbly deftand that the USSR withdraw Its troops froft Hungary?
MORRISON Q
Ves, because DEI'IAI'ID-4 Is TRUE.
SENTEI'ICE-269
DEMRI'ID-111
Question to ••k ARPA: (411D to end)
> Did the USSR withdraw its troops froft Hunesry?
MORRISON Q
Since WITHDRAW-2 Is a bracketed belief, I cannot be sure.
SENTENCE-278
WITHDRAW-11
Question to ask ARPA: (GilD to end)
> Uhat directive• ere there?
I'IORRISON Q
There are: REQUEST-1 REQUEST-2 CDI'InAI'ID-1 ~D-2 DEI'IAND-1 DEI'IRI'ID-4 DEI'IAI'ID-3 II'ISTRUCT-1
SEI'ITENCE-271
BE-2112
Question to aak ARPA: (GilD to end)
> ~net retractive con•teteftent• ere there?
I'IORRISOI'I Q
Does RETRACTIVE have .n ADVERB forft? (V or I'll l'lo.
There ere no COI'ISTATEI'IEI'IT-7s.
SENTENCE-272
BE-284
commands related to the lexical classifier. Since its implementation, the

belief-system examiner has proved an invaluable tool for inspecting and
verifying knowledge structures.
Discussion
After a text has been prepared for parsing, the "debugged" text parses
and references at a rate of 0.24 seconds for the average sentence, which
multiplies out to about 450 to 600 pages per hour. 24 A number of support
tools available in the RELATOS environment, such as the editor mode,
simplify and speed the text preparation process. Interestingly, the discipline
of converting text to the immediate reference model forces users to think
closely about what they themselves must do to understand the sentences.
This reflective process is a continuing source of insights into how language
works, suggesting how computers might model it. A competent RELATOS
user can process about ten pages of raw text from a new domain in a
working day. 25 As domain-relevant vocabulary and background knowledge
are developed, the daily amount of new text processed should increase and
converge to the time required to make the text meet the processing model.
Lexical Classification
Systems of Lexical Recognizers

Recognizer Organization. Lexical recognizers generally have associated
words and selection constraints designed to find the various lexical realizations
that instantiate the category. If a user defines a number of recognizers, it
is convenient to organize them into independent collections, or lexical
classification systems, and to organize the collections hierarchically (or
heterarchically). When lexical recognizers are organized hierarchically, two
types can be distinguished: Base categories generally form its leaves and
recognize instances, whereas the abstract categories, above them in the
hierarchy, have no tokens or selection constraints to recognize instances.
Instead, they have specialization, generalization, and equality relationships
to situate them taxonomically with respect to base categories or other
abstract categories. 26
Instance Classification. Although some commands in the RELATOS text
editor mode and the belief-system examiner simply find instances of categories
and view their source sentences, as seen in Figure 15.5, it is generally more
useful to label the category instances. For this reason, every lexical recognizer
should have an associated concept reference specification. This makes
possible commands in the belief-system examiner and editor to lexically
classify instances. Lexical classification makes available the categorizations
Figure 15.5: The source sentences that refer to speech acts in the 1956 Soviet intervention in Hungary. The numbered verbs denote the
semantic node for the speech act to which the accompanying sentence refers. Multiple appearances of the same numbered verb reflects
intersentential references to the same relation by different sentences.
OEftRM0-1: Th~ f"'ass~s de,.anded that IMre-Nagy replace Gero because the ,.,asses IMFDRn-1: J,re-Nagy 1 nforned the UN' s genera 1 secretary that add1 t 1 on a 1
were po 1 i t ; c i 'eed end because Ger-o did not set i sfy the ru11sse!ll and because the Soviet troops had entered Hungary.
""sses believed that the USSR had legitiP"'eted net tonal Con,unisn.
ACCEPT -1: The r~esses be 1 i eved that tl"le USSR had 1 egi t i noted net ton a 1 INSTRUCT-I: Inre-Nagy' s governMent requested that the UN-Secur i ty-Counci 1
COP'IMunisl"'' because the USSR accepted Go,..ulka in Poland. instruct the USSR to negotiate its diffel"'ences uith Inre-Nagy's goverm'lent.
COftftAMD-1: Gero uas ab 1e to cor~nend the secret police because the secret REQUEST-2: lnre-Negy' s govern,ent requested that the UN-Secur i ty-Counc i 1
police supported hin. instruct the USSR to negotiate its differences with J,re-Nagy 's governMent.
COftftltMD-2: The secret po 11 ce f t red so,.,e bu 1 1 ets at the !'lasses becauae the REOUEST-2: The USSR deadlocked the UN-Security-Counci 1 because the USSR d~d
"asses deMonstrated for I"re-Nagy and because Gero co,,.,anded the secret police not want the UN-Secur1 ty-Counci 1 to consider I"re-Nagy • s govern,ent 's request
to fire the bullets at the nesses. and because the USSR was able to deadlock the UN-Seeurity-Council.
COftftRMD-2:Gero cor,'landed the secret police to fire the bullets at the l""'asses REOUEST-2: The UN-Security-Counci 1 did not consider J,re-Nagy' 5 govern,ent' 5
because Gero feared that he uould fall fro" political pot.Jer. request because the USSR deadlocked the UN-Securi ty-Counc11.
PROIIISE-1: IP'Ire-Nagy pro,..ised to follou a Hungarian road-to-socialisP'I. REQUEST-2: J,..re-Nagy' s gover-nf'\ent' s request uas noved to the
UN-General-Assenb 1 y because the USSR deadlocked the UN-Secur i ty-Counc i 1 .
CRLL-1: R Hungarian politburo ne,..ber called for direct Soviet ,11 it.sry aid. DEnRND-3: The UH-General-Assenbly adopted a resolution ~.Jhich de,ande:d that
the USSR i""ediate:1y withdra~.J the soviet troops: fro,.. Huf'lgary.
RNMOUMCE-1: I "re-Nagy announced that Hungary ~o.~ou 1 d est ab 1 ish a f'IU 1t 1-party INfORn-2: Janos-l<adar' 5 governMent i nforf'led the- UN Genera 1-Secretary that
politic~l =y:=~ten in Hungary. Inre:-Nagy'5 conf'lunication5 ue:re: invalid.
REOUEST-1: France, the UK, and the USA reQuested that the: UN-Securtty-Counci 1 DE"AH0-4: The: UN-General-Assef'\bly continued to def'laf'ld that the USSR withdrau
consider Hungary's situation vis-e-vis the Soviet intervention. its troops fron Hungary although it ~oJBS clear that Soviet policy would not be
deflected and that the Soviet intervention had put down the Hungarian revolt.
CLRIII-1: France, the UK, and the USA claiP"'ed thet foreign ni 1 ttary forces CONCLUDE-I: lhe special co,..,ittee concluded that the revolt had been
were repress~ng the l"'ass:es• rights. spontaneous and nationalistic.
PROTEST-I: Inre-Nagy' s governnent prote5ted the UN-Secur 1 ty-Counc i 1 • s

consideration beeeuse the situation was within Hungary's do,..estfc juri5dict1on.
Figure 15.6: The LISP definition for a hierarchical lexical classifier that finds
the perlocutionary force of speech acts. lt looks for "cause" relations and
motivational "for" relations whose subject is a speech act. lt depends on the
prior lexical classification of speech acts.
(nEFINE-LEXICAL-CATEGORY
Pau.ocuriOIIART -FCIRCI:
"Find~ rel•tion.s cau.sed by .speech •ct.s. •
:RELATION
:LEXICAL-TOKENS (CAUSE f'OR)
: CONCEPT-REf'-SPEC
(REf'-SPEC
RELATION
:CONSTRAINTS
( (MINIMAL-INITIAL-DESCRIPTION
((UNIVERSAL-PI
(SUBJECT-RELATION HO PERLOCUTIONARY ( (TRUE) l)
CPNUMBER-OF-SUBJECT-RELATIONS HO 1) ) l)
:MODE
:REFERENCE)
:SELECTION-CONSTRAINTS
( (* ( (SUBJECT-CONSTRAINTS
((RELATION-PI
(CLASS ( (REf'-SPEC
ACT
:CONSTRAINTS
((MINIMAL-INITIAL-DESCRIPTION
((SUBJECT-RELATION Of'
(REf'-SPEC SPEECH
:QUANTIFICATION :TOKEN-UNIVERSAL
:MODE :INTERLEAVE)
( (PMSUBJECT-RELATION HO TIGHTLY-BOUND) l) l l)
:QUANTIFICATION :UNIVERSAL
:MODE :INTERLEAVE) l l l)
(OBJECT-CONSTRAINTS ( (RELATION-PI l l
(RELATION-PIll)
:LEXICAL-TYPE-STRING "Per locut ionary Force"
:LEXICAL-CLASSIFICATION-SYSTEM "Standard Lexical Classifications")
of a text model for other uses. These include displaying concept instances
or answering questions.
Hierarchical Classification. Hierarchical lexical classification uses prior
classifications in subsequent ones. For example, after recognizing speech
acts, another lexical classifier (Figure 15.6) could find perlocutionary force
(effects on others) of speech acts. Figure 15.7 illustrates one such recognition,
where the Soviet acceptance ("accept-1 ") of Gomulka in Poland caused the
Hungarian masses to believe ("believe-1 ") it had legitimated ("legitimate-1 ")
national communism. Figure 15.7 also shows that causal links continue
from the masses' beliefs to their demands ("demand-1") for lmre Nagy to
replace ("replace-1 ") Gero-reinforced by their politicization ("politicize-1 ")
and their dissatisfaction (" satisfy-1 ") with Gero. Instead of incorporating
speech acts into every lexical recognizer that needs to test if a relation is
a speech act, abstraction and modularity are best served by maintaining
separate classification systems, simply running them in the order of their
dependence, if unidirectional, or repeatedly to quiescence, if interdependent.
Although tractable in small applications, larger applications cannot afford
the overhead of irrelevant checks for category instances unnecessary for
the hierarchical recognition. Hierarchical lexical classifiers address this prob-
lem (Mallery, 1988c). They are just like ordinary base categories except
366 John C. Ma/Jery
Figure 15.7: The source sentence for a perlocutionary act and the semantic structure
lexically classified as the perlocutionary force in the 1956 Soviet intervention in Hungary.
The second graph shows the causal structure following from the masses' belief (believe-
1) that the Soviet Union had legitimated national communism to their demand (demand-
1) for Gero's replacement by lmre Nagy.
PERLOCUTIONARY-FORCE IRELATION-ll sentences in ARPA:
CAUSE-15: The masses believed that the USSR had legit1mated national
Communism because the USSR accepted Gomulka in Poland.
Instances of RELATION-1 In ARPA.
that they use classifications by other lexical recognizers in their recognition

constraints. The implementation records the dependencies on prior recog-
nitions and ensures those recognizers run first. The easy way to think of
hierarchical recognizers is to consider the base categories as the first layer
of recognizers and the hierarchical recognizers as subsequent layers. The
importance of hierarchical recognizers is that they allow composition of
complex patterns from smaller patterns, which each recognize some coherent
part. There are benefits; component categories are recognized along with
composite categories, and recognizers for composite categories are easier
to debug. Thus, a complex frame or script could be implemented bottom-
up with hierarchical recognizers that detect slots or roles and, finally, a top-
level recognizer that recognizes its instantiation in the pattern of recognized
slots.
Inspecting Results. A user can inspect the results of lexical classification
using the standard tools for examining semantic structure. The belief-system
examiner has a number of commands and display modes to view lexical
classifiers, their constraints, and the structures they pick out in text models.
lt is also possible to retrieve the source sentences for semantic structures
to see if classifications conform to usages in the source text (Figure 15.5).
Figure 15.8: The LISP definition of the lexical classifier for retractive speech acts
created using the definition interface shown in Figures 15.9 and 15.10. Before the advent
of the interactive editing interface, the user would have specified this definition directly
in LISP.
(DEFINE-LEXICAL-CATEGORY
UTRACriVI:-COIIS~ADIIEIIT
"'Find:J retr•ct1ve con•tatement.s. •
:VERB
:GENERALIZATIONS (:CONSTATEMENT)
:LEXICAL-TOKENS
(ABJURE CORRECT DENY DISAVOW DISCLAIM DISOWN RECANT RENOUNCE REPUDIATE RETRACT TAKE WITHDRAW)
:SELECTION-CONSTRAINTS
((TAKE
( ((SUBJECT-RELATION HAS-PARTICLE BACK ((TRUE)))
(OBJECT-CONSTRAINTS
C(OR ( ( CRELATION-Pl l
((CLASS CREF-SPEC PROPOSITION
:QUANTIFICATION :TOKEN-UNIVERSAL))))))))))
(WITHDRAW
( (OBJECT-CONSTRAINTS
(COR ( ( (RELATION-P))
( (CLASS ( (REF-SPEC PROPOSITION
:QUANTIFICATION :TOKEN-UNIVERSAL
:MODE :REFERENCE) ) ) ) ) )
CRELATION-P)) ) ) ) )
: CONCEPT-REF-SPEC
(REF-SPEC CONSTATEMENT
:CONSTRAINTS ((MINIMAL-INITIAL-DESCRIPTION
( (UNIVERSAL-P)
!SUBJECT-RELATION HO RETRACTIVE I (TRUE) l l
(PNUHBER-OF-SUBJECT-RELATIONS HO 1))))
:MODE :REFERENCE)
:OPAOUE-P T
:LEXICAL-TYPE-STRING "'Retractive Constatement"
:LEXICAL-CLASSIFICATION-SYSTEM "Bach i Harnish Speech Acts•)
The RELATUS editor mode supports many of the same commands but
does not use graphical displays for presentation. Both the belief-system
examiner and the editor allow a user to browse through semantic structure
using a mouse-sensitive "frame" inspector. Although question answering
based on classifications can provide another means of inspecting results
(Figure 15.4), people can more effectively look at the knowledge representation
than formulate questions to test classification correctness. In general, all
the implemented inspection methods rely on a human who examines, either
directly or indirectly, the classifications and the reasons for the classifications.
Defining Lexical Classifiers. Earlier versions of the RELATUS lexical
classification system (Mallery, 1987, 1988b, 1988c) required the analyst to
define categories by invoking a LISP definition form and supplying constraint
specifications (see Figure 15.8). Classifier definition required a rudimentary
knowledge of LISP and some familiarity with the constraint language. In
general, the ability of users to write pattern specifications limited the
complexity of lexical classifiers and recognitions. The recent introduction
of an editing interface for lexical classifiers pushed back these limits. The
editing interface
• automatically writes LISP definitions for lexical classifiers based on

information elicited from users via an advanced window interface;
368 John C. Ma/Jery
• automatically creates pattern descriptions in the constraint language

based on a series of English sentences provided by users.
Figures 15.9 and 15.10 show an example of editing a lexical classifier for
a speech act. There are three phases to specify each lexical realization's
constraint description.
• Specification Sentences: The user provides sentences to characterize

it (Figure 15.9).
• Specification Representation: The sentences are syntactically parsed
and semantically represented to yield a referentially integrated graph
structure.
• Constraint Description Construction: The semantic inverter traverses
the associated semantic representation and creates a constraint de-
scription to recognize similar knowledge structures (Figure 15.10).
Generalizing recognition constraints from single examples is a simple

form of explanation-based learning (Mitchell et al., 1986). 27
In the typical editing cycle, a user defines the newly edited lexical classifier
and then tests it, using invocation interfaces in the editor mode or the belief-
system examiner. If the lexical classifier retrieves instances that are not
members of the category (incorrectness) or fails to retrieve instances of
the category (incompleteness), the user may edit the lexical classifier to
add or remove constraining information. Once recognitions are adequate,
the individual definitions can be collected in LISP files that define a lexical
classification system. Thereafter, reloading the compiled versions of these
files reinstantiates the entire lexical classification system.
Existing Lexical Classification Systems

To date, there are six lexical classification systems.
• Lasswell Value Dictionary: The Lasswell Value Dictionary (Stone et al.,

1966; Lasswell and Namenwirth, 1969; Namenwirth and Weber, 1987)
was converted into a lexical classification system in October 1987. The
categories were organized taxonomically, but they are too extensive for
inclusion here. Practitioners of traditional computerized content analysis
expended considerable effort to develop this 8,000-word dictionary and
to verify the stability of its categories. 28
• General Problem Solving: A general lexical classification system for
belief, intention, and affect words has been developed primarily as an
example but was used to examine strategic language, including a political
Figure 15.9: Adding realization 12 to a lexical classifier for retractive speech acts by providing English sentences for
conversion into constraints. The resulting constraints will find and disambiguate the sense of verb "withdraw" that
means to retract a proposition.
0 MOR RISON 0 : Ed i t le ~<i cel Class i fier (New cl assifier [defau l t tio)) No
Le ""; eel categor y [defaul t Ret r a c t.i v e-Constater~ent] : Retract ive-Constate"ent
[dlt.lne Lexical Chsslfhr
L e xi cal Categor y: RETRACTIVE-CONS TATEMENT

Ooc ur~entat i on: Finds r•tract iv• eonst•t•,•nts .
Gr aph Cetegor y: Relation Objec t Token-Uni v ersal Universal Individual Verb Adjective Adv erb
Genera li Rat 1 ons : Constatenent
S pec i al i Eati o ns : k• yt.~ords or • null v•lu•
EQu i v alencl!s: k• yt.~or ds o r • null v a l u•
OpaQ u e categor y : Yes No
Le }( i ea 1 Ty pe : Retract iv e Constate l"' en t
L e ~<ic t~ lC l as!!l i f i cation Sy ste" : Bach & Harn i sh Spee ch Ac ts
Conc ept Ref-Spec : ( REF-SPEC CONSTATEMENT
'C ONSTRAIN TS
( ( "IN! MA L - !NIT I AL -DE SCRIPT I ON
(( UNI VERSAL-P ) (SUBJECT-RELATION HO RETRACTIVE ((TRUE))) ( PNU"BER - OF-SUBJECT-RELATIONS HO! ) ) ))
' "ODE
'REFERENCE )
L ~ ~dcel Real haUon 1: (ABJURE )
Le )o( ical Realil:etion 2 : (CORRECT )
Le i'( ical Realhation 3: (DENY )
Le Jod cal Rea l hat i on 4 : (DISAVOW )
LeJodcal RealhaUon 5: (DISCLAU1 )
LeMical Realh:ation 6 : (DISOWN )
Le Jod eel Real haU on 7 : (RECANT)
L e Jo~ ical Re:alhat1on 8: (RENOUNCE)
L e Jo~i cal Realh:ation 9: (REPUDIATE)
le ~'< ical Real hat ion 19: (RETRACT)
Le l'd eel Real hat 1on 11 : ( TAKE ( ( (SUBJECT -RELATION HAS-PARTICLE BACK ( (TRUE) ) )
(OBJECT -CONSTRAINTS
( ( OR ( (RELATION-P))
( (CLASS ( REF-SPEC PROPOSITION 'OUANTIFICATION ' TOKEN-UNIVERSAL)))))))) )
Le xi cal Rea 1 hat 1 on 12:
( Le l'(i cal token ) loiiTHDRAiol
( S~nt•nc•.s) A person withdraws the statel"'ent . The state~ue~nt ts a proposition .
( Ob j ect var iablh:ation le v el ) 1
( To ken constraints ) a const r aint $tructur•
<Mli!I> aborts , ~ uses tt'lese ~o.~a 1 ues
Dynamic Li&p Listener 1

~
99,161'89 9-4 : 48 : 48 JCI1a CL RL , !.l0....
Figure 15.10: The edited lexical classifier for retractive speech acts. The classifier now has a constraint description in w
lexical realization 12 to pick out the sense of "withdraw" as a speech act and a concept ref~spec to label category ~
instances.
0 MORRISON Q : Edit Le >dc el ClessHier (New classifier [default No]) No
I•Le"i eel category (default Retract i ve-Constate~u!nt): Retract 1 ve-Constete,..ent
Edttt"• leatcat Classtfler

L el'lic el Category: RETRACTI VE-C ONSTATE11ENT
Docunentetion: fihds r•tractiv• constat•,•nts.
Gr aph Category: Relation Object Token-Universal Universal lndh.ddual Verb Adjective Adverb
General i Eat 1 ons: Constate,ent
Sp e ci al i aations : Jc.yt.~ords or a null valu•
Equivelences: k•yvords or a nu ll valu•
Op aque categor y: Yes No
Le l'l i ce 1 Type: Retractive Consteter\ent
L e )'(i c~iil ClassHication Syste": Bach l Harnhh Speech Acts
Conc ept Ref-Spec: (R EF -SPE C CONSTATEMENT
:CONSTRAINTS
( ( I'll Nir,AL-INIT IAL-DESCRIPT ION \
((UNIVE.SAL-P) (SU8JECT-•ELATION HO •ET.ACTIVE ((T.UE))) (PNU"8E•-OF-SU8JECT-•ELATIONS HO I )))
'"ODE
'•EFE•ENCE)
l ~ "' i e ol R~olhot1on 1: (ABJURE)
l ~ 101ie el R~eliration 2, cco••EcT >
l ~ 101ie el Rea11aat ion 3, ( DENV)
l ~ "' e e l R~alhat ion •, ( DISAVO")
l ~ ~'~ e el Rea11rat 1 on 5' ( DISCLAI" )
l e ~'~ eel Reali a: at 1 on 6 ' ( DISO"N )
L e ~'~ eel Real i zat 1 on 1' ( •ECANT)
l~101ieel Reel hat 1 on 8' ( •ENOUNCE)
l~Jdeel Reeli zat 1 on 9 , (•EPUDIATE)
l~"'ieel Real i a: et ; on IS, (.ET.ACT)
l e "'ie el Reel het ; on 11 , (TAKE ( ( (SUBJECT -•ELATI ON HAS-PA.T I CLE SACK ( ( T•uE)))
(OBJECT -CONS! •AINTS
( ( o• ( ( ( •ELAT ION-P)) ( (CLASS ( •EF-SPEC P•OPOSITION , QUANTIFICATION ' TOKEN-U"IVE.SAL))))))))))
l ~ "'ie ol R~ol1rot.1on 12 : (IJITHDRAIJ ((OBJECT-CONSTRAINTS
((0. (((.ELATION-P))
((CLASS ((•EF-SPEC P.OPOSITION , QUANTIFICATION , TOKEN-UNIVE.SAL '"ODE '"EFE.ENCE))))))
(.ELATION-P)))))
l~"' i c al Reelhet ion 13: • l•xic•l r••liz•tion
<M2I> aborts, ~ uses these values
Dynamic Lisp Li& tener 1
99.11 b.I B9 94 : 59 : 19 keyboerd CL •L ' .!..!:..!_

actor's model of an adversary's (or an ally's) model of themselves

(M.allery, 1987).
• Sequential Prisoner's Dilemma: Roger Hurwitz developed a lexical
classification system for SPD protocols that identifies categories as-
sociated with conflict and social order formation (Hurwitz and M.allery,
1989; Hurwitz, 1990).
• Bach and Harnish Speech Acts: A lexical classification system has
been created for the Bach and Harnish (1979) taxonomy of speech
acts. Although the authors did not provide sense-disambiguation in-
formation, the addition of some selection constraints partially corrects
this drawback. 29
• Ortony Affect Lexicon: A lexical classification system has been created
for the Ortony Affective Lexicon (Ortony et al., 1987). lt distinguishes
various affect classes, such as internal versus external affect and
behavioral versus cognitive affect. This system also suffers from the
absence of selectional constraints.
• SHERFACS Actions: A lexical classification system was defined for
application to narrative precis automatically generated (M.allery, 1988b)
from the SHERFACS International Conflict Management Dataset (Sher-
man, 1987a, 1987b, 1988). This system categorizes the various actions
that a political actor may take in conflicts as coded in SHERFACS.
These action categories include the COPDAB categories (Azar, 1982).30
Analytical Applications
Semantic Content Analysis. The immediate political-analytic application
of lexical classification is semantic content analysis. The tools presented
above open a universe of ways to analyze texts, leaving behind many
problems of traditional computerized content analysis but bringing some
new ones. This section anticipates some evaluational issues for the meth-
odology.
Traditional content analysis has already faced the issues of reliability and
validity. 31 lt evaluates results by considering the reliability of a model and
the validity of the components of the model. For semantic content analysis,
reliability primarily concerns the text model and has several aspects:
• Stability: Stability refers to the temporal stability of preparing text for

machine parsing. When the same coder prepares input text at different
times, the resulting text model may reflect several different coding
practices within the processing model. Stability should encompass
c/assiflcational stability, or the effects on classification of variations in
coding practices. Instability can arise as coders make texts literal and
372 John C. Mallery
explicit. Allowing multiple word senses within parts of speech reduces

the changes to the text and therefore should enhance stability.
• lntercoder Reliability: Although stability measures the consistency of
private understandings, intercoder reliability, or reproducibility, measures
the similarities and differences between different coders. Ambiguous or
inconsistent coding rules for rendering text explicit can diminish re-
producibility. Similarly, cognitive factors, such as situation framing or
abstraction mismatches, may yield different codings. As in stability,
the concern here is with the impact of coding on the possible clas-
sifications.
• Accuracy: Accuracy of coding is the extent to which preparing text
for machine parsing conforms to the implemented processing model.
For example, expecting a "literal and explicit" system to resolve
metaphors would be an inaccurate coding. Similarly, exceeding gram-
matical coverage could lead to spurious syntactic analyses.
Semantic content analysis, as conceived here, should perform well in terms

of stability and intercoder reliability precisely because it reduces the amount
of coding humans must do.
Validity concerns the design of lexical classifiers and the interpretation
of results. Some considerations are:
• Construct Validity: Construct validity refers to the correlation of different

predictors for the same category instance. Two types of correlation
merit attention:
• Convergent constructs use different classifiers to identify the same
underlying concept;
• Divergent constructs use mutually exclusive classifiers, checking to
ensure that only one is present.
In semantic content analysis, construct validity can span both the coding
process and lexical classification. By retaining different surface statements,
lexical interpretive semantics avoids overloading semantically canonical en-
codings-an inherent problem for semantic universalism. Thus, good coding
practice seeks to retain alternate realizations of concepts while extending
lexical classifiers to identify them.
• Hypothesis Validity: Hypothesis validity is the extent to which the text

models yield classifications conforming to the substantive theory and
the procedural theory that the analyst purports to test. For example,
given a theory of organizational decision making and text models for
a specific organization, the semantic content analysis should produce
classifications consistent with both. To the extent that it does not,
there may be validity problems for the text model or the classifiers or
the theory of decision making (assuming the accuracy of the text
model and classifiers). Hypothesis validity raises the difficult problems
of establishing correspondence between the model and the external
world.
• Predictive Validity: Predictive validity is the extent to which the text
models yield classifications consistent with the phenomenal world. One
might compare the results of semantic content analysis against the
classifications of humans in order to establish the validity of the coding
for the text model as well as the lexical classifiers applied to it.
Since semantic content analysis depends on the reliable operation of

complex software systems and no complex computer system is ever bug-
free, it is desirable to find multiple derivations for results in order to reduce
the likelihood of spurious conclusions. Where this is not possible, the analyst
should verify the critical path leading to the result.
Beyond coding issues, different text types may have important conse-
quences for the validity of lexical classifiers (Hurwitz and Mallery, 1987).
Propaganda or insincere texts can produce spurious results for classifiers
devised for sincere texts. Insincere texts make inferential demands beyond
the processing model and would produce spurious classifications due to
inaccurate coding. Even if unanticipated sources of erroneous classifications
remain, at least text modeling allows explicit, reproducible representation
and analysis of texts.
Precedent Logics. The various types of precedential reasoning, such as
precedents repeated over time, analogies, and metaphor are distinguished
by the inductive distance between the target situation and the source situation
(Mallery and Hurwitz, 1987). For an AI system to retrieve source situations
for a given target, it is necessary to make a match specification that
generalizes the target situation. The two main heuristics for symbolic
generalization are:
• Climb Class Hierarchies to generalize relations and objects

• Drop Nonessential Relations to allow a match to succeed
Category information is crucial for an AI system to use either of these two

heuristics. Thus, the richer the classification of target and source situations,
the greater the likelihood of connecting these through some inductive
deformation of the target. Conceived as a search process, incremental
generalizations of the target find ever more inductively distant precedents
and analogies. Consequently, the quality of precedent search increases as
the classification of semantic memory becomes richer and more fine-grained.
374 John C. Mallery
Modeling Differential Perception. Lexical classification opens the possibility

of formally modeling differential classifications or "differential perceptions"
(Jervis, 1976). Here is how it could work. First, assume no convergence in
the lexical classification systems of actors and construct them independently.
This will help prevent bias in favor of convergence. The lexical classification
system can then be applied to the same texts, preferably with different
background knowledge and in different belief systems. The resulting clas-
sifications may then be examined to locate regions of convergent and
divergent classifications at the same or different levels of abstraction. In
addition to direct comparisons of classificational behavior, it is also possible
to investigate consequences for other facets of political cognition, such as
precedent search. Precedent retrieval depends on inductive deformations of
cases and on classification for assignment of mappings. If different clas-
sification systems are in use by political actors, the divergences in classi-
fications may lead to the retrieval of different precedents (even assuming
the same historical record). This may, in turn, lead to differential problem
framing.
Recognizing Argument Connectives. Computational argument analysis can
provide a formal basis for regrounding political science (Aiker, 1988a). Minsky
(1987) suggests that arguments can be treated like proofs, except that the
logical chain operates at a higher level of abstraction and makes larger, less
justified steps. Lexical classification could recognize argument connectives
if they were explicit. Unfortunately, most precomputational theories of
argument seem to be At-complete, or require arbitrary inferences to identify
the connective relationships. Largely syntactic theories emphasizing explicit
lexical markers-possibly Rescher's (1977) dialogical theory of argument,
which has shown political relevance (Alker, 1988b)-could be usefully
recognized. This kind of application for lexical classification shades into
parsing semantic structures and, in principle, could even extend to complete
deductive inference (McAIIester and Givan, 1989).
Conclusion
Lexical classification for text models yields the new methodology of
semantic content analysis. Analysts can use this method to rigorously and
reproducibly simulate classifications in political texts and political action.
For international politics, the method can support studies of how convergent
and divergent classification figure in conflict and cooperation. Its hermeneutic
grounding in interpretive semantics anticipates differential interpretation as
it insulates against distortions originating from the modeling tool itself. Since
classification begins from an eidetic representation grounded in the words
and grammatical relations of the original text, it insulates against analyst
bias; though the analyst may overlay his theoretical vocabulary on the
phenomena, that vocabulary does not provide the ultimate ground of the
text model. This increases the validity of analyses.
More generally, text modeling provides a new representational foundation
to formal models in political science. Because this foundation is ontologically
and epistemologically neutral, it can support culturally, ideologically, and
politically neutral analysis. Future research may extend text modeling from
recognition and generation of arguments to a new symbolically grounded
decision science. As extensions of text modeling come to support evolutionary,
cognitively informed world system models, the emerging social science
workbench may come to be an indispensable research associate for political
scientists.
Notes
1. This chapter was improved by comments from Carl Hewitt, Valerie M. Hudson,
Robert P. Weber, and Jeremy M. Wertheimer. The implementation follows a path
pioneered by Boris Katz and Patrick Winston. Gavan Duffy's parser helped make
this research possible. The encoding of syntactic parses builds from the Winston-
Katz research and evolved from debates with Duffy about the correct division of
labor between syntax and its description in logical form. Analysis of game protocols
with Roger Hurwitz since 1985 motivated the development of the lexical classification
system reported here. The moral support and foresight of Hayward R. Alker, Jr.,
allowed imagination to become reality. Marvin Minsky, Carl Hewitt, Berthold Horn,
Gerald Sussman, and countless past and present members of the MIT Artificial
Intelligence Laboratory made up the unique discursive background that situates this
research. Lincoln Bloomfield encouraged sober applications to practical problems.
Any shortcomings are the sole responsibility of the author.
The author was partially supported by a National Science Foundation Presidential
Young Investigator Award number DDM-8957464 to D. Sriram, Department of Civil
Engineering, MIT. Some earlier research was partially funded by a John D. and
Catherine R. MacArthur Foundation grant for research on international security and
arms control to the MIT Center for International Studies. This chapter describes
research done at the Artificial Intelligence Laboratory of the Massachusetts Institute
of Technology. Support for the MIT Artificial Intelligence Laboratory's artificial
intelligence research is provided in part by the Advanced Research Projects Agency
of the U.S. Department of Defense under Office of Naval Research contract number
N00014-85-K-0124. RELATUS research has also been done at American Microsystems,
Inc., and Symbolics, Inc.
2. Computerized content analysis deploys a form of keyword search to find the
frequencies and correlations of interesting words. Word frequency is deemed to
reflect salience and, in turn, importance for a speaker. Word senses are disambiguated
according to word eo-occurrences. But standard content analysis programs neither
analyze syntactic structures nor construct referentially integrated semantic repre-
sentations. Holsti (1968), Krippendorf (1980), and Weber (1985) provide overviews
376 John C. Mallery
of content analysis. Duffy (1987, 1989) explains how computational hermeneutics

transcends traditional content analysis.
3. Lexical recognizers store constraint descriptions for different paraphrases, or
alternate lexical realizations in surface semantics. Recognizers can incorporate selection
constraints so their constraint descriptions will disambiguate words senses, picking
out correct instances. Category instances are found as these constraint descriptions
match knowledge structures in the text model. Lexical recognizers do not attach a
tag word to sentence strings like conventional content analysis. Instead, each recognizer
stores a separate constraint description that denotes the knowledge structure for
the concept it recognizes. As recognizers locate instances, they label them with a
subsumption relation from the knowledge structure for the concept to the instance
knowledge structure. As lexical recognizers find and label instances, explicit taxonomic
structures grow up from the instances toward more general and abstract categories.
4. The RELATUS Natural Language Environment consists of 5.1 megabytes of
Common LISP source code, to which each implementor holds individual copyrights,
1983-1990. Duffy created the syntactic parsing system (1.7 megabytes). Mallery
created the semantic components (3.4), including the GNOSCERE system, reference
system, semantic inverter, lexical classification system, question answerer, belief-
system examiner, and RELATUS text editor. Although Mallery created the sentential
constraint poster (.3), its semantic encoding of parses was a collaborative effort.
Mallery also created the feature vector editor system (.8), which is used with Duffy's
Timebase system (.25).
5. Knowledge representation schemes based on formal logic (Konolige, 1984;
Moore, 1985) require the analysis of belief operators before a specific representation
can be defined and therefore selected. Thus, these formalisms cannot represent
natural language until comprehensive logical analyses are available.
6. Mallery, Hurwitz, and Duffy (1987) introduce the hermeneutics literature relevant
for artificial intelligence.
7. In practice, individual histories are not available, so analysts must reconstruct
in a psychobiographical mode relevant histories as information resources allow.
Earlier forms of cognitive modeling in political science, such as cognitive mapping,
faced a similar difficulty.
8. By analogy to NP-completeness in complexity theory, Fanya S. Montalvo coined
the term "AI-complete" to denote a computational problem whose difficulty is
equivalent to solving the central AI problem, i.e., making computers as intelligent
as people.
9. Winston (1980, 1984) and Katz (1980; Katz and Winston, 1982) used Katz's
deep structure transformational parser to build representations in Winston's tenary-
relation frame system. RELATUS is a descendant of that system but differs from
it in important ways discussed in this paper
10. RELATUS has used the Symbolics Flavor System extensively for object-
oriented programming. The implementation will be revised as the new standard for
LISP, the Common LISP Object System (De Michael, 1989), becomes available.
11. Since natural languages allow us to talk about practically anything, including
this sentence, the safest approach is to use an arbitrarily expressive representation
and restrict its expressibility only as necessary for specific applications.
12. While working with the Winston-Katz system in 1980-1983, the author could
not represent texts larger than about live pages on early LISP machines because
of "page thrash." Thrashing occurs when a computer with a virtual memory
architecture (real memory and swapping to disks) cannot swap in the task into real
memory because the task elements are two widely scattered in memory or too large
for real memory. One of the initial motivations for beginning work on RELAT(JS
during the summer of 1983 was to develop a knowledge representation system that
could avoid thrashing yet represent enough text for nontrivial social scientific
applications.
13. The intern function returns the symbol object given its print name and
package.
14. The reference system does not backtrack during the constraint application
process. The analog of backtracking is the number of unsuccessful possibilities that
appear in the initial possibility space.
15. This discussion draws from earlier descriptions by Duffy (Duffy and Mallery,
1986: 21, 22; Duffy, 1987).
16. Because sentential look-ahead is unbounded, the parser is not deterministic
in the sense of Marcus (1980). However, because operations terminate at clausal
boundaries, the parser is effectively deterministic. It remains a polynomial LR(k,t)
algorithm, although k is variable, not constant. See Berwick and Weinberg (1984:
192) for a discussion of the time complexity of LR(k,t) parsers.
17. Katz (1980; Katz and Winston, 1982) applied this computational principle in
his parser, demonstrating that deep-structure transformational parsing was tractable
and thereby refuting the intractability argument against deep structure (Winograd,
1971: 197).
18. Actually, the syntactic configurations, grammatical relations, carry forward
into the semantic representation through the ref-specs. But decoding goes directly
back to deep structure because the generation direction does not pose a graph-
matching problem.
19. The analysis in Duffy and Mattery (1984) suggests that this might not be
easy for all linguistic theories.
20. The deictic context of text is a collection of indexicals associated with a
text that includes, for example, the source, the recipient, the coding location, the
reference location, and the coding time. These indexicals situate the text and allow
the reader to resolve references for pronouns such as "me," "you," or indexical
nouns such as "here," "there," and "now." This facility constitutes the outside of
a context mechanism. It was designed according to categories suggested by Levinson
(1983: 54-94).
21. The text was coded in literal and explicit English from the account in
Butterworth and Scranton (1976) between 1980 and 1983. The author originally
coded the text for input to the Winston-Katz analogy system. The author successfully
performed some analogical reasoning with this text but was never able to fully
represent its 6,000 nodes until the advent of the representational technologies in
RELATUS.
22. The sentence generator handles multiple clauses, relative clauses, as well as
passive and dative transformations but lacks numerous stylistic and morphological
378 John C. Mallery
capabilities. The generator creates a theory-independent base structure, then uses

a theory-specific module to decode semantic structures derived from the Duffy
parser. Mallery (1989) presents general principles that make a complete semantic
inverter and a complete constraint-interpreting reference system possible.
23. Due to the irregularity and the popularity of copular questions (e.g., "Who
is Joe?"), a special facility handles them.
24. These timings are for the new Symbolics XL 1200 LISP machine with 4
megawords of real memory.
25. Attaining this level of competence may require several months of study. The
ten page/ day figure applies to the older Symbolics 3600 series LISP machines,
which run at one-sixth the speed of the XL 1200.
26. The implementation allows base categories to have specializations.
27. Although the semantic inverter supports relation variabilization, it has not
yet been incorporated into this interface.
28. Because the string-oriented disambiguation rules of the Lasswell Value Dic-
tionary have not been converted to RELATUS-style, grammatically based constraints,
the fruit of this earlier work have not yet been fully gleaned.
29. The lexicon appears as an appendix to Mallery (1987).
30. Mattery (1988c) reports on these classifiers.
31. This section draws on Weber (1985: 16-21) and Krippendorf (1980: 130-
154) but reformulates their criteria for semantic content analysis.
Bibliography
Alker, H. R., Jr. (1975) "Polimetrics: Its Descriptive Foundations." In F. Greenstein
and N. Polsby, eds., Handbook of Political Science, vol. 7. Reading, MA: Addison-
Wesley, 1975: 140-210.
___ . (1984) "Historical Argumentation and Statistical Inference: Towards More
Appropriate Logics for Historical Research." Historical Methods. Summer 1984,
17 (3): 164-173.
___ . (1985) "Reflective Resolution of Sequential Prisoner's Dilemmas." Paper
presented at the 1985 Annual Meeting of the Society for General Systems Research.
March 1985.
_ _ . (1986) "Bit Flows, Rewrites, Social Talk: Toward More Adequate Informational
Ontologies." Proceedings of the Tokyo University Conference on Information and
Its Functions. 1988.
___ . (1988a) "Politics as Political Argumentation: Towards a Regrounding of the
Global Discipline." Paper presented at the XIVth World Congress of the International
Political Science Association. Washington, D.C., August 1988.
_ _ . (1988b) "The Dialectical Logic of Thucydides' Melian Dialogue." American
Political Science Review. 1988, 82 (3): 805-820.
Alker, H. R., Jr., Bennett, J., and Mefford, D. (1980) "Generalized Precedent Logics
for Resolving Insecurity Dilemmas." International Interactions. 1980, 7: 165-206.
Alker, H. R., Jr., and Christensen, C. (1972) "From Causal Modeling to Artificial
Intelligence: The Evolution of a UN Peace-Making Simulation." In J. LaPonce and
P. Smoker, eds., Experimentation and Simulation in Political Science. Toronto:

University of Toronto Press, 1972.
Alker, H. R., Jr., and Hurwitz, R. (1980) Resolving Prisoner's Dilemmas. Washington,
D.C.: American Political Science Association, 1980.
Alker, H. R., Jr., Hurwitz R., Mallery, J. C., and Sherman, F. L. (1989) "SHERFACS
Data and the Butterworth Narrative for the Caribbean/Cuban Missile Crisis in
the Light of New Evidence: More Powerful Textual Modeling Possibilities and
Changing Scholarly Purposes." Paper presented at the 1989 Joint Annual Convention
of the British International Studies Association and the International Studies
Association. London, England, March 1989.
Alker, H. R., Jr., and Tanaka, A. (1981) "Resolutional Possibilities in 'Historical'
Prisoner's Dilemmas." Paper presented at the 1981 Annual Meeting of the
International Studies Association. Philadelphia, March 1981.
Axelrod, R. (1984) The Evolution of Cooperation. New York: Basic Books, 1984.
Azar, E. E. (1982) The Codebook of the Conflict and Peace Data Bank. College Park:
Center for International Development, University of Maryland, 1982.
Bach, K., and Harnish, R. M. (1979) Linguistic Communication and Speech Acts.
Cambridge: MIT Press, 1979.
Barsalou, L. W., and Bower, G. H. (1984) "Discrimination Nets as Psychological
Models." Cognitive Science. 1984, 8: 1-26.
Barwise, J., and Perry, J.(1983) Situations and Attitudes. Cambridge: MIT Press,
1983.
Bennett, J. P. (1987) "Deterrence in Contemporary Policy Debates: The Expression
of Strategic Doctrines in Political Arguments about SDI." Paper presented at the
1987 Meeting of the International Studies Association. Washington, D. C., April
1987.
Berwick, R. C., and Weinberg, A. S. (1984) The Grammatical Basis of Linguistic
Performance. Cambridge: MIT Press, 1984.
Butterworth, R. L., with Scranton, M. E. (1976) Managing Interstate Conflict, 1945-
74. New York: Knopf, 1976.
Butterworth, R. L. (1980) Managing Interstate Conflict, 1945-79. State College:
Department of Political Science, Pennsylvania State University, final report to
ARPA, February 1980.
Chapman, D. (1985) "Planning for Conjunctive Goals." Artificial Intelligence. July
1987, 32: 333-377.
Charniak, E., and McDermott, D. (1985) Introduction to Artificial Intelligence. Reading,
MA: Addison-WP.sley, 1985.
Chomsky, N. (1965) Aspects of a Theory of Syntax. Cambridge: MIT Press, 1965.
Dallmayr, F. R. (1984) Language and Politics: Why Does Language Matter to Political
Philosophy? Notre Dame: University of Notre Dame Press, 1984.
De Michael, L. G. (1989) "Overview: The Common Lisp Object System." Lisp and
Symbolic Computation. January 1989, 1 (3/4): 227-244.
Devereux, E. A. (1989) "Computational Analysis of Similarities and Differences in
Textual Descriptions: A Comparison of Newspaper Coverage of the 1967 Election
in South Vietnam." Paper presented at the 1989 Annual Meeting of the American
Political Science Association. Atlanta, August 1989.
380 John C. Mallery
Deutsch, K. W. (1953) Nationalism and Social Communication: An Inquiry into the

Foundations of Nationality. Second edition, Cambridge: MIT Press, 1966. The first
edition was originally published in 1953.
Duffy, G. (1986) "Categorial Disambiguation." AAAI-86: Proceedings of the American
Association for Artificial Intelligence. August 1986: 1057-1061.
- - . (1987) Language, Politics, and Method: Grounding a Computational Her-
meneutic. Cambridge: Doctoral dissertation, Department of Political Science, MIT,
August 1987.
- - . (1988) "Theory and Text: How Textual Analysis Can Help Build International
Relations Theory." Paper presented at the Annual Meeting of the International
Studies Association. St. Louis, March 1988.
- - . (1989) "The Validity Problem of Content Analysis." Paper presented at the
Annual Meeting of the American Political Science Association. Atlanta, August
1989. This is a revised chapter from Duffy, 1987: 21-40.
Duffy, G., and Mallery, J. C. (1984) "Referential Determinism and Computational
Efficiency: Posting Constraints from Deep Structure." AAAI-84: Proceedings of
the 1984 National Conference on Artificial Intelligence. August 1984: 101-105.
- - . (1986) "RELATOS: An Artificial Intelligence Tool for Natural Language
Modeling." Paper presented at the 1986 Meeting of the International Studies
Association. Anaheim, March 1986.
Feigenbaum, E. A., and Simon, H. A. (1984) "EPAM-Iike Models of Recognition and
Learning." Cognitive Science. 1984, 8: 305-336.
Fodor, J. A., Garrett, M. F., Walker, E.C.T., and Parkes, C. H. (1980) "Against
Definitions." Cognition. 1980, 8: 263-367.
Gentner, D., and Landers, R. (1985) "Analogical Reminding: A Good Match is Hard
to Find." Paper presented to the panel on "Commonsense Reasoning" at the
International Conference on Systems, Man and Cybernetics. Tucson, Arizona,
November 1985.
George, A. L. (1959) Propaganda Analysis: A Study of Inferences Made from Nazi
Propaganda in World War 11. White Plains, NY: Row, Peterson, 1959.
Grice, H. P. (1975) "Logic and Conversion." In P. Cole, and J. L. Morgan, eds.,
Syntax and Semantics 3: Speech Acts. New York: Academic Press, 1975: 41-
58.
_ _ . (1978) "Further Notes on Logic and Conversation." In P. Cole, ed., Syntax
and Semantics 9: Pragmatics. New York: Academic Press, 1978: 113-128.
Habermas, J. (1979) "What is (]niversal Pragmatics?" in J. Habermas, Communication
and the Evolution of Society. T. McCarthy, trans., Boston: Beacon Press, 1979.
_ _ . (1981) The Theory of Communicative Action: Reason and the Rationalization
of Society. Thomas McCarthy, trans., Boston: Beacon Press, 1981. German edition
published in 1981.
Harre, R. (1985) Personal Being. Cambridge: Harvard (]niversity Press, 1985.
Hobbs, J. A. (1979) "Coherence and Coreference." Cognitive Science. 1979, 3: 67-
90.
_ _ . (1986) "On The Coherence and Structure of Discourse." In L. Polanyi, ed.,
The Structure of Discourse. Norwood, N.J.: Ablex, 1986.
Holsti, 0. (1968) "Content Analysis." In G. Lindzey and E. Aronson, eds., The
Handbook of Social Psychology, volume 2. Reading, MA: Addison-Wesley, 1968.
Hurwitz, R. (1990) Re: Solving the Prisoner's Dilemma. Cambridge: Doctoral disser-
tation, Department of Political Science, MIT, 1990.
Hurwitz, R., and Mallery, J. C. (1987) "RELATUS Users Notes." Cambridge: MIT
Artificial Intelligence Laboratory, draft, 1987.
___ . (1989) "Norms in Sequential Prisoner's Dilemma Protocols: Using New
Techniques of Semantic Content Analysis to Identify Normative Orientations in
Computer Models of Text." Paper presented at the 1989 Annual Meeting of the
American Political Science Association. Atlanta, August 1989.
Hurwitz, R., Mallery, J. C., Alker, H. R., Jr., and Duffy, G. (1986) "Anarchy or
Community: A Study of Developmental Patterns in SPD Games." Paper presented
at the 1986 Meeting of the International Studies Association. Anaheim, March
1986.
Jackendoff, R. S. (1972) Semantic Interpretation in Generative Grammar. Cambridge:
MIT Press, 1972.
Jervis, R. (1976) Perception and Misperception in International Politics. Princeton:
Princeton University Press, 1976.
Lakoff, G., and Johnson, M. (1980) Metaphors We Live By. Chicago: University of
Chicago Press, 1980.
Katz, B. (1980) "A Three Step Procedure for Language Generation." Cambridge: MIT
Artificial Intelligence Laboratory AI Memo No. 599, 1980.
_ _ . (1988) "Text Processing with the START Natural Language System." In
E. Barrett, ed., Text, Context, and HyperText: Writing with and for the Computer.
Cambridge: MIT Press, 1988: 55-77. An earlier version was published as "Using
English for Indexing and Retrieving." Proceedings of the Conference on User-
oriented Content-based Text and Image Handling, RIAO 'BB. A longer version
appears in MIT Artificial Intelligence Laboratory AI Memo No. 1096, October
1988.
Katz, B., and Winston, P. H. (1982) "A Two-way Natural Language Interface." In P.
Degano and E. Sandewall, eds., Integrated Interactive Computing Systems. Amer-
sterdam: North-Holland, 1982.
Katz, J. J., and Fodor, J. A. (1963) "The Structure of a Semantic Theory." Language.
1963, 39 (2): 170-210.
Konolige, K. (1984) A Deduction Model of Belief and Its Logics. Menlo Park: Artificial
Intelligence Center, SRI International, Technical Note 326, 1984.
Krippendorf, K. (1980) Content Analysis: An Introduction to its Methodology. Beverly
Hills: Sage, 1980.
Lasswell, H. D. (1927) Propaganda Technique in World War I. Cambridge: MIT Press,
1971.
- - . (1935) World Politics and Persona/Insecurity. New York: Free Press, 1965.
Lasswell, H. D., Leites, N., et al. (1949) Language of Politics: Studies in Quantitative
Semantics. Cambridge: MIT Press, 1949.
Lasswell, H. D., and Namenwirth, J. Z. (1969) The Lasswe/1 Value Dictionary. New
Haven: Yale University Press, 1969.
Lenat, D. B., and Guha, R. V. (1990) Building Large Knowledge Based Systems:
Representation and Inference in the Cyc Project. Reading, MA: Addison-Wesley,
1990.
382 John C. Mallery
Levinson, S. C. (1983) Pragmatics. Cambridge: Cambridge University Press, 1983.

Lyons, J. (1977) Semantics, volume 1, Cambridge: Cambridge University Press, 1977.
McAIIester, D., and Givan, R. (1989) "Taxonomic Syntax for First Order Inference."
Cambridge: MIT Artificial Intelligence Laboratory AI Memo No. 1134, June 1989.
Maida, A. S., and Shapiro, S. C. (1982) "lntensional Concepts in Propositional
Semantic Networks." Cognitive Science. 1982, 6: 291-330.
Mallery, J. C. (1985) "Universality and Individuality: The Interaction of Noun Phrase
Determiners in Copular Clauses." Proceedings of the Association for Computational
Linguistics. 1985.
- - - . (1987) "Computing Strategic Language: Natural Language Models of Belief
and Intention." Paper presented at the 1987 Meeting of the International Studies
Association. April 1987.
- - - . (1988a) "Thinking About Foreign Policy: Finding an Appropriate Role for
Artificially Intelligent Computers." Cambridge: Master's thesis, MIT Department
of Political Science, MIT, 1988. A revised version was presented at the 1988
Meeting of the International Studies Association. April 1988.
___ . (1988b), "Beyond Correlation: Reconstructing the Relational Structure of
Conflict Processes from SHERFACS Data." Paper presented at the 1988 Meeting
of the International Studies Association. April 1988.
___ . (1988c) "Identifying Implicit Categories in Text Models: A Step in Com-
putational Hermeneutics." Paper presented at the XIVth World Congress of the
International Political Science Association. Washington, D.C., August 1988.
___ . (1989) "The Invertibility Principle of Knowledge Representation." Cambridge:
MIT Artificial Intelligence Laboratory, draft.
___ . (1990) "Constraint-Interpreting Reference." Cambridge: MIT Artificial In-
telligence Laboratory AI Memo No. 827, forthcoming. A draft version of this
memo has seen limited circulation since March 1985.
Mallery, J. C., and Duffy, G. (1990) "A Computational Model of Semantic Perception."
Cambridge: MIT Artificial Intelligence Laboratory AI Memo No. 799, forthcoming.
A draft version of this memo has seen limited circulation since January 1986.
Mallery, J. C., and Hurwitz, R. (1987) "Analogy and Precedent in Strategic Decision-
Making: A Computational Approach." Paper presented at the Annual Meeting of
the American Political Science Association. September 1987.
Mallery, J. C., Hurwitz, R., and Duffy, G. (1987) "Hermeneutics." In S. C. Shapiro,
The Encyclopedia of Artificial Intelligence. New York: John Wiley, 1987: 362-
376.
Mallery, J. C., Hurwitz, R., with Alker, H. R., Jr., and Duffy, G. (1986) "Analyzing
Natural Language Protocols of Sequential PD Games." Paper presented to the
Microcomputer Group at the 1986 Meeting of the American Political Science
Association. Washington, D.C., August 1986.
Malone, T. W., Grant, K. R., Lai, K. R., Rao, R., and Rosenblit, D. A. (1987) "Semi-
Structured Messages are Surprisingly Useful for Computer-Supported Coordina-
tion." ACM Transactions on Office Information Systems. 1987, 5: 115-131.
Marcus, M. P. (1980) A Theory of Syntactic Recognition for Natural Language.
Cambridge: MIT Press, 1980.
Mefford, D. (1987) "Analogical Reasoning and the Definition of the Situation: Back
to Snyder for Concepts and Forward to Artificial Intelligence for Method." In C.
F. Hermann, C. W. Kegley, Jr., and J. R. Rosenau, eds., New Directions in the

Study of Foreign Policy. Boston: Alien & Unwin, 1987: 221-244.
_ _ . (1988a) "Women, Fire, and Artificial Intelligence." Paper presented at the
colloquium "Words, Deeds, and Social Things." Lincare College, Oxford University,
March 11-13, 1988.
_ _ . (1988b) "An Alternative to Rational Choice: The Function of Stories in the
Formulation of Foreign Policy with Examples from U.S. Interventions." Columbus:
Department of Political Science, The Ohio State University, unpublished paper,
summer 1988.
Michalski, R. S. (1987) "Concept Learning." In S. C. Shapiro, The Encyclopedia of
Artificial Intelligence. New York: John Wiley, 1987: 185-194.
Mitchell, T. M., Keller, R. M., and Kedar-Cabelli, S. T. (1986) "Explanation-Based
Generalization: A Unifying View." Machine Learning. 1986, 1: 47-80.
Minsky, M. (1987) The Society of Mind. New York: Simon and Schuster, 1987.
Moore, R. C. (1985) " A Formal Theory of Knowledge and Action." Palo Alto: Center
for Study of Language and Information, Stanford University, report no. CSLI-85-
31, 1985.
Namenwirth, J. Z., and Weber, R. P. (1987) Culture Indicators Research: A Content
Analytic Perspective. Boston: Unwin Hyman, 1987.
Newell, A.(1980) "Physical Symbol Systems." Cognitive Science. 1980, 4: 135-183.
_ _ . (1989) The 1987 William James Lectures: Unified Theories of Cognition.
Cambridge: Harvard University Press, 1989.
Ortony, A., Clore, G. L., and Foss, M. A. (1987) "The Referential Structure of the
Affective Lexicon." Cognitive Science. July-September 1987, 11(3): 341-364.
Peirce, C. S. (1901) "Abduction and Induction." Reprinted in J. Buchler, Philosophical
Writings of Peirce. New York: Dover Publications, 1955. Written circa 1901.
Piaget, J. (1979) The Child's Conception of the World. J. Tomlinson and A. Tomlinson,
trans., Ottawa: Littlefteld Adams, 1979.
Pool, I. S., Lasswell, H. D., Lerner, D., et al. (1970). The Prestige Press: A Comparative
Study of Political Symbols. Cambridge: MIT Press, 1970.
Rescher, N. (1977) Dialectics: A Controversy-Oriented Approach to the Theory of
Knowledge. Albany: State University of New York Press, 1977.
_ _ . (1978) Peirce's Philosophy of Science: Critical Studies in His Theory of
Induction and Scientific Method. Notre Dame: University of Notre Dame Press,
1978.
Ricoeur, P. (1971) "The Model of Text: Meaningful Action Considered as Text."
Social Research. 1971, 38: 529-62. Reprinted in J. B. Thompson, ed. and trans.,
Paul Ricoeur: Hermeneutics and the Human Sciences. Cambridge: Cambridge
University Press, 1981.
- - . (1975) "Phenomenology and Hermeneutics." Translated and reprinted in J.
B. Thompson, ed. and trans., Paul Ricoeur: Hermeneutics and the Human Sciences.
Cambridge: Cambridge University Press, 1981. Originally published as "Phenom-
enologie et Hermeneutique." Phamenologische Forschungen, vol. 1, E. W. Orth,
ed., Freiburg: Kart Alber, 1975: 31-77.
- - . (1977) "Metaphor and the Semantics of Discourse." In The Rule of Metaphor:
Multi-Disciplinary Studies in Metaphor. R. Czerny et al., trans., Toronto: University
of Toronto Press, 1977: 65-100.
384 John C. Mallery
Russell, B. (1948) Human Knowledge: Its Scope and Limits. New York: Simon and
Schuster, 1948.
Schank, R. C. (1972) "Conceptual Dependency: A Theory of Natural Language."
Cognitive Psychology. 3, 552-631 (1972).
Schank, R. C., and Abelson, R. P. (1977) Scripts, Plans, Goals, and Understanding.
Hillsdale, N.J.: Lawrence Erlbaum, 1977.
Searle, J. R. (1969) Speech Acts. Cambridge: Cambridge University Press, 1969.
_ _ . (1979) "Metaphor." In Andrew Ortony, ed., Metaphor and Thought. Cambridge:
Cambridge University Press, 1979: 92-123.
Shapiro, M. J. (1984) Language and Politics. New York: New York University Press,
1984.
___ . (1988) The Politics of Representation. Madison: University of Wisconsin
Press, 1988.
Sherman, F. L. (1987a) Partway to Peace: The United Nations and the Road to
Nowhere. State College: Doctoral dissertation, Department of Political Science,
Pennsylvania State University, 1987.
___ . (1987b) "Four Major Traditions of Historical Events Research: A Brief
Comparison." Paper presented to the MIT /DDIR Conference on "New Directions
for Storing, Indexing, Retrieving, Coding and Analyzing Information on International
Events." Cambridge: Center for International Studies, MIT, November 1987.
___ . (1988) "SHERFACS: A New Cross-Paradigm, International Conflict Dataset."
Paper presented at the 1988 Meeting of the International Studies Association. St.
Louis, Missouri, March 1988.
Simon, H. A. (1962) "The Architecture of Complexity." Proceedings of the American
Philosophical Society. December 1962, 106: 467-482. Reprinted in H. A. Simon,
The Sciences of the Artificial. Second edition, Cambridge: MIT Press, 1981.
Originally published in 1969.
___ . (1985) "Human Nature in Politics: The Dialogue of Psychology with Political
Science." The American Political Science Review. 1985, 79: 293-304.
Stone, P. J., Dunphy, D. C., Smith, M. S., Ogilvie, D. M., with associates (1966)
The General Inquirer: A Computer Approach to Content Analysis. Cambridge:
MIT Press, 1966.
Symbolics, Inc. (1988) Reference Manual for Symbolics Computers. Cambridge:
Symbolics, Inc., 1986. Revised 1987, 1988.
Taylor, C. (1985) Human Agency and Language: Philosophical Papers. Cambridge:
Cambridge University Press, 1985.
Waltz, D. L. (1975) Understanding Line Drawings of Scenes with Shadows. In P. H.
Winston, ed., The Psychology of Computer Vision. New York: McGraw-Hill, 1975:
19-91.
Weber, R. P. (1985) Basic Content Analysis. Beverly Hills: Sage, 1985.
Winograd, T. (1971) Procedures as a Representation for Data in a Computer Program
for Understanding Natural Language. Cambridge: Doctoral dissertation, Department
of Electrical Engineering and Computer Science, MIT, 1971.
Winograd, T., and Flores, F. (1986) Understanding Computers and Cognition: A New
Foundation for Design. Norwood, NJ: Ablex, 1986.
Winston, P. H. (1980) "Learning and Reasoning by Analogy." Communications of

the ACM. December 1980, 23(12).
_ _ . (1984) Artificial Intelligence. Reading, MA: Addison-Wesley, 1984.
Wittgenstein, L. (1953) Philosophical Investigations. New York: Macmillan, 1958. First
edition published in 1953.
16
Time Space: Representing Historical Time
for Efficient Event Retrieval
GavanDuffy
At the risk of stating the profoundly obvious, time is an essential aspect

of politics. Political events do not occur randomly. They occur within orderly
sequences of events-political processes. Articulating the general features
of those processes constitutes a primary task for students of politics.
Because intentional human actors interpose themselves between the
political explanans and the political explananda, methods successfully applied
to similar tasks in the natural sciences find much more limited applicability
in political explanation. Adequate empirical political theories must contain
propositions that are conceptual as well as contingent. That is, while
accounting for observed regularities in "objective" variables, they must also
account for the practical inferences of actors, whenever their deliberative
actions in response to stimuli-independent variables-produced observed
changes in dependent variables (Moon, 1975).
These practical inferences may rely on a wide variety of premises drawn
from the sociopolitical contexts within which political action takes place.
In institutional contexts, for example, formal rules or opposition strength
may limit the action possibilities of particular decision-makers. Likewise,
particular cultural contexts can constrain the bounds of legitimate political
action, as interpreted subjectively.
Deliberative political choices are also, in part, a function of the times in
which they are made. Can contemporary British policy concerning the
"troubles" in Northern Ireland be understood without reference to the
exigencies of European integration? Can European integration be understood
without reference to changing patterns of international trade? Can the
withdrawal of the Soviets from Afghanistan, the defunding of the Nicaraguan
Contras, or the simultaneous pullback of Cuban troops from Angola and
South African troops from Namibia be understood without reference to the
thaw in East-West relations? And can the thaw in East-West relations be
386
Time Space: Representing Historical Time 387
understood without reference to fiscal crises in both the Soviet Union and
the United States? Certainly the answers to all these questions must be
negative, and resoundingly so.
In explaining the political decisions that produce political events, then,
political analysts must remain mindful of the temporal (and other) contexts
within which they occur. This chapter describes a computational tool designed
to assist analysts in maintaining and analyzing the temporal contexts of
events. 1 Its computational substrate has recently been extended to incorporate
geographic as well as temporal contexts. Timebase is a program that indexes
the time intervals corresponding to the durations of events data. The indexation
scheme supports the efficient retrieval of those data. Initial motivation for
Timebase sprang from the RELATUS text-analysis system, discussed else-
where in this volume. The next section details this motivation. A following
section reviews the standard AI technique for representing temporal intervals,
revealing its central weaknesses. lt also describes our reconceptualization
of the temporal domain, developed simultaneously by Rit (1986). Subsequent
sections describe in lay terms the computational substrate of Timebase and
present Timebase in operation using the data in Frank Sherman's SHERFACS
international conflict data set (Sherman, 1985, 1988), represented in John
Mallery's Feature Vector Editor (Mallery and Sherman, 1988). Finally, a
conclusion discusses extensions to Timebase and possible political-analytic
applications of its computational substrate.
The Motive for Temporal Indexation

The primary need for Timebase follows from our need in RELATUS to
represent and reason about time. Time can be expressed in natural language
by means of tense (Reichenbach, 1947: 287-298; Hornstein, 1977, 1981;
Comrie, 1985; Yip, 1986), aspect (Woisetschlaeger, 1976; Comrie, 1976),
and "temporal adverbials"-adverbs, prepositional phrases, and clauses that
temporally restrict the referents of verb phrases. To illustrate, suppose the
following sentences were represented in a RELATUS belief system:
• From 1929 to 1933, Henry L. Stimson was the secretary of state.

• From 1933 until after World War 11 had begun, Cordell Hull was the
secretary of state.
• On January 7, 1932, the secretary of state declared that the United
States did not intend to recognize any agreement that might impair
the sovereignty, the independence, or the territorial and administrative
integrity of the Republic of China.
Now suppose we queried RELATUS, "Who declared that the United States
does not intend to recognize any agreement that might impair the sovereignty
388 Gavan Duffy
Table 16.1: Thirteen Possible Temporal Relations

NAME INVERSE RELATIONSHIP
Superior Inferior xxxxxxxxxx

yyyy
Same Start Same Start xxxxxxxxxx
Superior Inferior yyyyyyy
Same End Same End xxxxxxxxxx
Superior Inferior yyyyyyy
Overlap Under lap xxxxxxx
yyyyyyy
Equality Equality xxxxxxxxxx
yyyyyyyyyy
Starts at End Ends at Start XX XXX
yyyyy
Disjoint Before Disjoint After xxxx
yyyy
of a country?" 2 Without some means of representing the distinct temporal

contexts within which Stimson and Hull served as secretary of state, the
system becomes hopelessly confused. At best, it can only answer that
either, or worse, both Stimson and Hull made this declaration. But this is
clearly wrong. Stimson made the declaration, and since that information is
present in the text, the system should be able to answer the query correctly.
Timebase does not exhaust the needed machinery for reasoning about
time, however. It is designed to serve as a back-end temporal representation.
It will prove necessary but not sufficient for temporal reasoning. In addition
to Timebase, several components are needed. Ordinal time relationships,
specified in the semantic representation, are also necessary, as are procedures
for updating these relationships and the interval relationships in Timebase.
Temporal Representation
The Pointer Propagation Problem

Pointer propagation is a central problem for implementors of AI temporal
representation schemes. The difficulty can be illustrated first by considering
A.
B.
c.
D.
E.
Time
Figure 16.1: Intervals Incapable of Hierarchical Expression
the possible temporal relationships in Table 16.1 and then considering the
line segments in Figure 16.1.
In Table 16.1, the Xs and Vs represent two intervals along the time line.
The table exhaustively specifies the thirteen possibl~ temporal relations
identified by Alien (1984). The pointer propagation problem surfaces when
one attempts to create a network of relations in which each interval points
to its immediate neighbors in the during hierarchy, as well as to its immediate
overlap and underlap. For example, consider Figure 16.1. Here, each line
segment represents a time interval along the unidimensional time line. The
left-hand end point represents the start time of the interval while the right-
handed point represents its end time.
Overlaps and underlaps prohibit hierarchical expression of interval relations.
Interval A is temporally superior to intervals B, C, D, and E. However, B
is neither inferior nor superior to C, and E cannot be expressed as related
to C or D in any hierarchy of temporal superiority /inferiority. In general,
for any pair of intervals related along this superiority /inferiority dimension,
there are an infinite number of possible intervals that are inferior to the
superior interval but neither inferior to nor superior to the inferior interval.
So, unless one is willing to lose track of some intervals completely, it makes
no sense to represent intervals in a temporal during hierarchy. The during
hierarchy is for this reason fictitious.
One way to attempt resolution of this difficulty would be to make all
intervals point to any interval to which it is related (except, of course,
disjoint intervals). Unfortunately, a procedure for retrieving all intervals that
are related in some way to a particular interval might cause an exhaustive
search through the entire data base of intervals. Any such operation would
be hopelessly slow. Worse, unless that procedure remembered each interval
it had already visited, there would be no guarantee that the search would
terminate. Alternatively, each interval could maintain a direct pointer to any
other interval to which it is related. Here, the representation would consume
an inordinate amount of space. This is the pointer propagation problem.
Alien hopes to finesse the difficulty by introducing reference intervals.
390 Gavan Duffy
Each time a relation is added, it is added with respect to a reference interval.

The user of the system may set up the reference intervals as desired. In
general, they probably will often reflect part of the during [superior-inferior]
hierarchy, but may also reflect the semantic clustering of events (Alien, 1981:
223).
This approach does not solve the problem. It only diminishes it. Reference
intervals may themselves be incapable of hierarchical expression. More
importantly, limiting search within particular "semantic clusters" may blind
the program to the eo-occurrence of events that are semantically distal but
temporally proximal. To relate this point to international relations, two inter-
state disputes may involve different sets of states, concern substantively
different issues, and occur in widely separate geographical locations. Never-
theless, they may occur at about the same time and their eo-occurrence
may not have been accidental. Alien's approach is incapable of detecting
such semantically disparate temporal eo-occurrences without exhaustive
search of the data base.
Reconceptualizing Time
The pointer propagation problem is solved by reconceptualizing time
itself. Instead of considering time intervals as segments of the time line,
consider them as points in a two-dimensional Cartesian coordinate system.
Let the X axis represent the start times of intervals and let the Y axis
represent the end times of intervals. Call this space time space. Consider
the origin, G, to represent the genesis point, or the beginning of all time.
Since no time interval can end before it starts and no part of any interval
can occur before the beginning of time, all possible time intervals could be
represented as points in the upper half of the upper right-hand quadrant
of the space. Let this be the temporal octant. Note that all zero-length
intervals fall on the diagonal that divides the quadrant. Let this be the
moment diagonal. Figure 16.2 presents these constructs graphically.
For any particular interval, the spatial location of all other intervals related
to it by any of the thirteen possible temporal relations in Table 16.1 is
completely defined by the locations of (a) the interval itself, (b) the start
axis, and (c) the moment diagonal. To illustrate, Figure 16.3 displays the
interval X in a close-up view of the temporal octant. In Figure 16.3, the
spatial locations of superiors, inferiors, overlaps, underlaps, and disjoints
before and after are shown relative to the interval X.
All superiors of any time interval are located in that portion of the
temporal octant higher in the end dimension and lower in the start dimension.
Conversely, all of its inferiors are located in the portion of the octant higher
in the start dimension and lower in the end dimension. These areas are
391
moments
I
start = = = = = = = = G = = = = = =
Figure 16.2: Time Space: A Coordinate System for Time
c D
I I disjoin Is
after
overlaps
superiors
I
I
B X
I E
under laps inferiors
A
F
disjoints
before
Figure 16.3: An Interval's Relations in the Temporal Octant

392 Gavan Duffy
Table 16.2: Temporal Relations Falling on Line Segments

RELATION SEGMENT
Same Start Superiors CX

Same Start Inferiors XF
Same End Superiors BX
Same End Inferiors XE
Starts at End DE
Ends at Start AF
defined by lines parallel to the start and end axes that run through the
point representing the interval. In Figure 16.3, these lines are BE and CF.
Points E and F lie on the moment diagonal. From E, we draw a line in
the temporal octant parallel to the end axis. 3 In Figure 16.3, this is line
DE. From F, we draw a line in the temporal octant parallel to the start
axis. This is line AF. All of the interval's underlaps are found within the
area bounded by AF, FX, BX, and the start axis. All of its overlaps are
found in the area bounded by CX, EX, and DE. This area extends upward
infinitely in the end dimension. All intervals that are disjoint before the
interval appear in the triangle bounded by AF, the moment diagonal, and
the start axis. Intervals disjoint after the interval appear in the area bounded
on two sides by DE and the moment diagonal and extending infinitely upward
in the end dimension.
Intervals that are equal to the interval of interest are represented as the
same point in the space (X in Figure 16.3). The other possible temporal
relations in Table 16.1 fall on line segments in Figure 16.3. These are listed
in Table 16.2.4
Implementing the Reconceptualization

The reconceptualization does not help unless efficient techniques are
available for storing and retrieving time interval data. Bentley's k-dimensional
binary search trees (Bentley, 1975; Bentley and Friedman, 1979; Bentley
and Maurer, 1980) were used for this purpose. Briefly, k-dimensional trees
are similar to binary search trees. They differ in that the dimension for
comparing nodes alternates at each successive level of depth in the tree.
In the two-dimensional temporal case, then, comparisons alternate between
415-817
~ ~316-942
42-382
15-~ ~404 28~ ~71

181L ~448 5~995
Figure 16.4: Balanced 20-Tree for Randomly Selected Intervals
the end times and start times of intervals. This method produces a tree
that progressively bifurcates the two-dimensional space.
For example, consider the small tree of intervals in Figure 16.4. All
intervals whose end times are greater than 817 are found to the right of
the root interval, and all less than 817 are found to its left. Similarly, intervals
whose start times precede the two intervals at the second level are found
to their left, while intervals whose start times follow those start times are
found to their right.
Our reconceptualization of time effectively reduces the problem of retrieving
intervals superior to, inferior to, overlapping, etc., any interval to a range
search in the two-dimensional time space. The two-dimensional (20) tree
representation supports this operation efficiently, and the pointer propagation
problem consequently disappears. The search procedure simply dives into
the tree from the root, proceeding only in directions that satisfy the range
search parameters. As it dives into the tree, it collects all intervals that
satisfy those parameters. When the tree is balanced, this happens in time
proportional to log2 N+ A, where A is the number of intervals that fall within
that range. 5
These logarithmic results are important. They mean that operations will
not be slowed dramatically as the size of the data base increases. This
efficiency depends critically, however, on balance. That is, operations will
remain efficient only when every node in the tree has roughly the same
number of progeny to its left as to its right. If this is not the case, storage
and access operations slow monotonically with increases in the size of the
data base. To ensure balance and thereby to ensure efficiency, the two-
dimensional trees must remain balanced at all times, regardless of the order
in which intervals are stored or retracted. Methods for maintaining balance
in two-dimensional trees were developed and subsequently extended to the
k-dimensional case. These procedures run in time proportional to N[Iog2 N+ 1)
in the worst case. However, its performance in the average case is much
more efficient.
394 Gavan Duffy
Timebase in Operation
The Timebase implementation can currently represent time intervals
starting anytime after January 1, 1000, and ending before January 1, 2900.
The level of resolution is the second. Presently, Timebase unrealistically
assumes a Gregorian calendar and daylight savings time during the appropriate
portion of each year throughout this 1,900-year period. The level of resolution
is arbitrary and can easily be altered for other applications. The points in
Figure 16.5 present the Timebase screen representation of 5,000 intervals
selected at random from July 4, 1776, to April 2, 1988.
By following the tree branching, the search procedure finds all the intervals
within any particular range on each of the two axes, start time and end
time. For example, to find all time inferiors of the interval X, the search
procedure is constrained to find all intervals that (a) start after X starts,
(b) start before X ends, (c) end after X starts, and (d) end before X ends.
Figure 16.5 shows the result of searching for the time inferiors of the
interval that starts at 7:54:27 p.m. on October 17, 1874 and ends at 7:42:24
p.m. on July 31, 1915. The search procedure uses the constraints of this
particular interval's range to walk the 20 tree only in the direction of
intervals that can satisfy those constraints.
As it walks the tree, the search procedure carries along another procedure
and applies it to every interval that satisfies the search constraints. In the
case of Figure 16.5, the procedure carried along simply prints out the
interval in the window in the bottom left-hand corner of the screen. However,
any arbitrary procedure may be applied to all intervals within any particular
range of start times and end times.
Figure 16.6 displays the timebase for 1,737 phase structures in the
SHERFACS data set. Note that, unlike the distribution in Figure 16.5, the
distribution of intervals in Figure 16.6 is far from random. Intervals tend
to cluster close to the moment diagonal, reflecting the fact that many phases
have relatively short life cycles. 6 Here is where the balance procedures help.
Figure 16.6 shows how the balanced tree partitions the intervals. Because
the tree is balanced, the partitions in the denser regions of the temporal
octant are smaller. Thus, search operations perform as fast in these denser
regions as they do in the sparser regions away from the moment diagonal.
To convey a feel for user interaction with Timebase, and to demonstrate
its practical utility for exploring the data, several screen snapshots of a
practice session were taken. In the session, I first used Mallery's Feature
Vector Editor to filter the SHERFACS data set (Mallery and Sherman, 1988),
constructing a set that included only those conflicts for which there was
a threat to superpower (U.S. and USSR) interests that could, in the eyes
of the superpower (as determined by the coders) change the global military
Figure 16.5: Search Space for Time Inferiors
[""
Time base ,....
Se lee\. l •me bate Scnw"1>1ot
Select J.,urvol Subocotwr
Rctct 1 •mcbatc [rou ''23'1 taB .. i .~.. . . . .. . . '
; . • I -~ • ., ..,
Pa,.tition Vec\..Of" ·.' .. ~ -~·· : _,
.
,. ,;
Show Tru Int.ei"T'' 1'2t,a7' , ~ ·,·. .,.., . ... ... ·.·. /
Oot.o Relotiono Rclat•ont
Oot.o Rceoon Litt Rce•on ' . :~
···:,
5/11'1"3 : ·~
OV.Cr (hr-oniclc
- . . ' ·:: ~.:... ..·.. /
'
tore of: ''14,1958
., ·.·.,· <."<.' .~·:~ -,:"
II,..r.,.
,..,,,,,.,. 19a,.a2'1 u ,.....,,,.,,, lta42 al4 . · -~ -~ .· ~· .
1'11'1938
.. .;.. : ...;,·'."
............, .. ..... u .,,,.,,,. • ., •• •• • ... ,• ·. . ~;; .. .. ;,
._,11;'1..,., IO : JI•I6 U J , J..,II'9t I'J : J3 : )o& .;,
5,9,1925 ~ - ~,. ' • ' I '
~··~ . ,,. ,,. u , ........ ,, , •• .,. ;;
~1'1- 12 a,.lll U 1,.....1192 16 : 12 : "
.... .,...., . ,. _,, ... ~ .... .,,, ,J, ,,4,1912 :·.·. ~ ..-
........... " •••at••t .......... .,. 11 ,., , 2'1
........1'1 . , , , , . . . . ,.......,,.., ,, , , , . .:.;• ,.. . .. ,. ··.
12,31,1199 ·'o:
......... ..,., , ~
,......,,.., G : n oJt u ,.......,..,. 11•111••1
'~·~
. . ............... .. ,..,.,
... ,. •• , u ,....,, ,..,., ., ,., ,,. 4.'29,1887
&,.1/11,. 1J1S21JI' \e..,,..,,., ll a, o2"
............, ., ,,.., ,, u ,, , ,, • ., 17•4l: Jt ,,25,1874
"''..,'""' ,.,,.,,, ....... ,..... ,.., ,,,,., .
1'1'1~ G••J •:U U ""''* U • ll =el 12'21 '1861
6;"2,,., •=,,,,. ... ,....,,,,., ......,.
.. ,.....,_ . , ... , ,. ... ~, ... ,.., ) 11 42 • 13 ~·
.,....,., . , , . . , ,, ... ,.... , .~ J h tl : 7l 4,19,1149
• ...,.... . . , . 1111'7 : ~ ... , , , , , . . " : JD : .
. · ...,."
.....,.,, .. .,. JlaZ2 ••' U ~'"' : SJ•e:J ''15,1836 I
. . -z:J,,.,. ... ,,,., ... .,....,,,,., I'...... ,
...., .. , .. . .. ...... . . , ... .... 1): 21•11
12,12,1823 '·:·'/
..,,.,,IT? Utll • l2 u ,..,._..,_, • =11•2'9 ./
......,,_ z:t , • •,. u ~ ..... , •• .,,S3
... · · ~;; ...
,,...... '" 2l • JI : s.J \.e .~ . . . ll : lel .. 4,11,1811
. .) -~ /
..,.,, ,_ n •• , ,,. u '~' " '"' .,,,.,,
,,,21' .........., u ,...,.....,_ Cl • Jl ··
..', ~ /
1-'5,1798
,....., .... 11•11: 2:2 u ,.......,,_ • . Je • ll
.. )
, ,, , . , 12 •1'l •t6 u , . , • • 112: . . . .
14'1 ... n ... ,, u ,...,....,.., ,,,., ,, , 12'2'1785 t "
IJJ"S..I_, 12•el••' U ,._,,.,,_ Z2 :SJ : :rt >
u ,.,.,,_ • =Sl•" 8'3i'1113 I ' ' ' ' t
........... ~ • • a : ll Y .....,.,.,,_ . , :I I •ZS
5,26,1757 120,1894 ,,19,1852 12,31'1899 ''14.'1947 1'24'19
I6A h"'onatr-attOf't ft,..ebaae 5ta,.t Tt"'c
I ,,.,,., .,,,.,.2
0
ltende,.d Dev1et 1 on
Mlect er'l 1nte,.vel with the II'IOutc. 49 ,...,.. 58 week•
18'11'1847 14 : 13 : 49
1 i "'ebetc C.Of'IP\end : I raa • 5tlt"t
...." 49 .,...,., 49 weeke 1 «My \.A)
11.,.ebe•c c~end : ••let iona ["" 11'25'1 91 7 12 : 21 : 14
the i,.,t.,....,., it : ,.,,,, ,,,. ,,,,.,2"1 u ,.,,,,,., , , , .,,~ Dut"etion 78 2 weckt
l'••'"• 49 ,..,.. 13 .... ~-::. ' dl)"t ~
Figure 16.6: SHERFACS Conflict Phases Partitioned to Depth 6
~
Timebase (nd ··~
Sc: l£cl luTitbut Scat~.o«rt>lo t
Sclccl )nl-0.-vol SubiCitt.tr
R(lft l•mcbasr [,...,, 5/38,1988
Part1t.10n Vcct.or
Sho"" Tree lnt.en'l 7'23,1985
Data Rr lation• Relat ions
Oat-a Rceion lilt Rr9ion ,,1,,1 982
Ot.N• (hror,.c lc
I 11 ,9,197!
1'1'1977
2/25,1974
4/28,1971 r . . 1. ·!]
''12'1968 . ... [7
r
1'6'1 965
,,29/1962
11'22'1959
1'15/1957 t
' I
31'111'195•
lll&fp
5/3/1951 ~ !~t
''26,1948
·~9~~5~ ],'·
18/12/1942
I I I I
12/6/1939 I '
.,,,1914 7,9,1928 18'12'1942 1/15,1957 4/28,1971 7/23'19 5
r
5fte,.F'ACS Star-t T1"'e
N • 1727 5tandard Deviation

T t ,.,.baac CO"'I"IInd i lttcr•"•nt Pertitto'l 11 ,....,., U 1111111eka 5 Nya
COt"'l"llnd
&tart ,,26'1964 . . : 26 : 25
""""
T i~baec i lfl~'•"•"t '•rtitlott 12 yeara
(nd 1/4/1967 12 : 58 : 48
T i '-boac COf'll"let'\d i lttero,.•nt '•rtttion .. .. ,..,.. 41 . . . k. 2 dlyl
Bur-at iol'l 3 ~,,., 5 week 1 4 daya
T il"lcbeer C'Of'l~"~l"'d I
balance between them. These phases were then mapped into another
Timebase, presented in Figure 16.7.
Notice that there are two relatively dense clusters in the time space. The
relatively sparse area that separates them seems to span start times that
occurred during the Johnson administration.
By selecting the chronicle command in the command pane (in the upper
left-hand portion of the figures), Timebase presents the user a set of menu
options. Selecting "Build New Chronicle," another menu appears prompting
for the temporal specifications of the chronicle. In order to chronicle the
conflict phases that started during the Johnson administration, the temporal
constraints "from November 22, 1963, to January 20, 1969," were specified.
Note that a range of end times and minimum and maximum durations could
also have been specified. Given an analytical interest in knowing what
conflict phases were initiated during this period, though, it made sense not
to constrain end times and durations.
Once the constraints are specified, the range search procedure walks the
2D tree in the direction of the region of time that the constraints denote. 7
Intervals that satisfy the constraints are given to a chronicle object. Once
the search terminates, selection of the chronicle menu item produces a
more elaborate chronicle menu.
Once the chronicle is built, the chronicle construction procedure asked
whether to display the chronicle graphically. An affirmative answer causes
the display in Figure 16.8 to be drawn. This represents each interval in the
chronicle as a line segment on the time line. There were six intervals in
the chronicle-three of relatively short duration that ended prior to the end
of the Johnson administration and three of longer duration that ended after
Johnson had left office.
By selecting the "Scatterplot" item on the command menu, the scatter
diagram reappears. Reinvoking the chronicle menu and selecting the option
"Draw Chronicle Region," identifies the region of the scatter diagram covered
by the chronicle, as indicated in Figure 16.9. This verifies that the Johnson
administration did indeed encompass the sparse region of the diagram
between the two dense clusters.
The option "Write Chronology" causes a description of the events (in
this case SHERFACS conflict phases) to be written to an editor buffer. The
buffer that was produced is presented in Table 16.3.
Analysis of the chronology revealed an interesting fact. All the conflict
phases in the chronology except one concerned the Southeast Asian region.
The exception was the settlement of the Aswan High Dam dispute, while
all the Southeast Asian phases were crises, hostilities, or post-hostility crises
or disputes. A number of possible explanations could be pursued, using this
discovered fact as a point of departure. The purpose of this session, which
Figure 16.7: Timebase of Threats to Superpower Military Balance
Time base [ncl 11-
~
s.roct l•mebeu SceLLc~loL
Select lnwrvel Sut>tceLLcr
Run l•mcbeoc [,..,. 11'21/1,.8
PV\•L•on Vector ,
~- T..ee lnw"' 3/12/1,.6 ,
O.te RclaLiOfta Rclat.•ont ,.
Oete Rc1i0t1 li•t Rc:eion ,,2,1tl3 /
I OV..r Chr-onicle ,.
I I 11'21"1918 ,.
/
2'11,1978 /
I'
/
6,3,1975 .,
/
,,22,1972 /
/
1,12,1t71 /
/
$/4/1t67 /
,
/
''23,1t64 /
/
I2/IV1t61 ,
,
.,5,1t59 . , !
,
,,25'1t56 7
ll/14/1953 ,
.,;,
3,6,1951
,
/
- . /
,,26,1948 /
,
ll/16/1945 •>'
/
2,5,1943
,,,,lt31 ''12,1!39 ,,1.'19417 6'J''I9~S 5,18,1963 4/28'1971 l/22,1979 2'22'19b
"'-'"'ec• To Superpower Jnte,.eata
,...,.,.c. Sta~"'t T 1 ...,
tt ...eba•• eo~~: Sutterplot ...... St.af'\Cflr"d hOJiet.ton

lh,.b. . e c-PICI: I ''-•~""t.
118 .,16,19~6 12:18:45 t ,..,... 11 weeka 6 cll•v•
[PICI 5'22,1t68 21:19:13 11 .,..,... 24 w.eke 1 day
Our-et ton 4 ~·"• 5 weeka 2 deva 6 ,..,.. 21 weeka 1 day
r.
Figure 16.8: Displaying Intervals in the Chronicle
Timebase
X lccl l1mtbau Scolw..,lol
~ l <cl J,wrvol Subocotwr
Ret~l Timebesr r.....
Pa,.t.it.ion Vcc\..Or
She"' Tre< lnw"'
0.1.41 Rolol•ono Rclat.iont
0.1.41 Roe•on li1t. Rteion
OV.or (t.,,.,;,,.
I
- ---
I
' ' ' ' '
,,23,198~ .. ,2 ..
;l,t
11'17'1961 ''"'I 966 4,21'19'1 1'21'1 , , 11'21'1981 fl,..
Chronicle in ""-'"'•'-• Threat To 5uoe,.oo.,.,. Jnt.e,.eata .
8v11d l'tew tn,.on1c.le

. llun 5tancla,.d 0,.~,~1ation
luilt . 0 •
lP&
5tlr"t 4,16,1956 12 : .. : 45 t ...,. • • ¥elkl ' . . ,...
a,,.dldine tn,.ontcle . 11 .,..,.. 24 ....k. 1 . .,..
Dte.olay it? (Y o,. ") Yea . End 5,22,1961 21 : 19 : 13
4 ,.. •• ,.. 5 .,..... 2 . . ,... 6 ~*'"' 21 weeka 1 day
'11'11e'biiC COI"'I"'If"'d : I
r.
O.,at ion ~
Figure 16.9: Drawing the Chronicle Region 8
[nd liN
Time base
Se le et T1mcbasc Scolt..o!'l>iol
~lrcl lnt..orvo l St.~btcatt.cr
Rctct Timc:bate En se 11'28'1988
Pa,..ti t.ion Vcct.or ,
Show Tree lnwm 1'12'1"6 /
Oot. l!rlotiont Relation•
Oot. l!eeion litt Rc~ i on U2'1"3
, ·,
OV.or Chronicle
Lo~at ''"'l : 11'22 '1 903 8e : 88 : 88118,21'19B8
,
H1fhcat 5url : 1'2e,l9b9 88 : 88 : 88 ,
I 2,18,1978 . I
/
,,3,1975 '/
,,22,1972 /
1'12'1978
5'4'1967
lt23t1964
I
12114'1961 .,
.,5,1959 .. . . I
7t25t1956
., ,
.. ,
/
11 '14/) 953 ,
,
lt6fl951
.,
6t2VI,.B
18/16'1945
,,· I
,
2t5t1943 I .< ' • ' t
919'1 931 1'12,1939 7t!<,l94 7 6/16'1955 5'18'1963 4/28,1971 1'22,1979 2'22'19 7
5M,..feca Threet To S~o~perpover Jntereata 5tart. TiN
,. • 178 5tenderd h~o~ i et; on

t i,...baae CON'\end : Chrcmitle Ste,.t .,1 ,,1956 82 : 18:45
"""" t ,..,., 11 weeka 6 deya
Drhl Ch,.onicle lceio,., 5,22'1968 28 : 19:13 11 ,_..,,.., 24 ... eka I o...,
Dcacribc Cur,.ent Chr-onicle Duration 4 vt•'"• 5 weeka 2 clava ' )l'e:lf"l 21 .,.,.k. 1 dl)l
T hwbeae CON"'end : '""
401
Table 16.3: The Chronology Constructed from the Chronicle

Temporal Constraints: Lowest Start: 11/22/1963 00:00:00
Highest Start: 1/20/1969 00:00:00
1/1/1964 00:01:00
Start: US-PRC Overflights, 1964-1975 (7)
Phase: Crisis
[Ends at 2/28/1971 23:59:00]
4/1/1964 00:01:00
Start: Laotian Civil War 12. 1964-1977 (21)
Phase: Hosti 1 ity
[Ends at 8/7/1975 23:59:00]
5/9/1964 00:01:00
Start: Aswan High Dam, 1955-1964 (11)
Phase: Settlement
[Ends at 12/31/1964 23:59:00]
12/31/1964 23:59:00
End: Aswan High Dam, 1955-1964 (11)
Phase: Settlement
[Started at 5/9/1964 00:01:00]
4/5/1966 00:01:00
Start: PRC-Bong Kong Tensions, 1949-1984 (17)
Phase: Hostility
[Ends at 12/31/1967 23:59:00]
12/1/1967 00:01:00
Start: PRC-Bong Kong Tensions, 1949-1984 (17)
Phase: Post-Hostility Crisis
[Ends at 12/31/1984 23:59:00]
12/31/1967 23:59:00
End: PRC-Bong Kong Tensions, 1949-1984 (17)
Phase: Hostility
[Started at 4/5/1966 00:01:00]
7/30/1968 00:01:00
Start: Malayan Independence, 1945-1968 (33)
Phase: Post-Hostility Dispute
[Ends at 12/31/1968 23:59:00)
12/31/1968 23:59:00
End: Malayan Independence, 1945-1968 (33)
Phase: Post-Hostility Dispute
[Started at 7/3~/1968 00:01:00]
2/28/1971 23:59:00
End: OS-PRC Overflights, 1964-1975 (7)
Phase: Crisis
[Started at 1/1/1964 00:01:00]
8/7/1975 23:59:00
End: Laotian Civil War 12, 1964-1977 (21)
Phase: Hostility
[Started at 4/1/1964 00:01:00]
12/31/1984 23:59:00
End: PRC-Bong Kong Tensions, 1949-1984 (17)
Phase: Post-Hostility Crisis
[Started at 12/1/1967 00:01:00]
402 Gavan Duffy
consumed less than five minutes, was to construct a miniature example of

the use of Timebase and its chronicling tools for event analysis.
Prospects
Along with its use as an indexation mechanism with which to ground
temporal inferences in the RELATUS text analysis system, Timebase could
serve as the critical temporal component of a general "history machine."
This history machine could be useful as both a teaching tool8 and-as
Bennett (1984) envisions-an analytic tool.
Although the Timebase representation has immediate utility for event
analysis, particularly when integrated with the Feature Vector Editor and
event data sets such as SHERFACS, there are certain limitations that could
be overcome. For example, the most finely grained time unit in SHERFACS
is the day, while Timebase uses the second as its time unit. 9 The temporal
codes of events data bases, which are often coarser than those in SHERFACS
and occasionally are even ordinal (e.g., Feierabend et al., 1972), drastically
limit the sorts of time-based analyses that might be performed.
With more finely grained temporal codes, one might begin to analyze
the temporal substructure of events. Such analyses might examine coup
contagions and terror contagions, 10 effects of transnational communication
lags, temporal constraints on mobilizing military forces, the relative delays
for achieving domestic consensus for political action across regime types
and transnational consensus for collective action across alliances, etc.
Attention to temporal detail in data creation efforts, then, could open
opportunities for generalizing about the details of international event processes.
This sort of processual analysis presupposes application of modern pattern-
matching capabilities, such as those Mallery has implemented in RELATUS.
Unfortunately, existing event data sets too often "flatten" the structure of
events in order to build data matrices. Data flattening rips events from their
semantic contexts, radically reducing opportunities for pattern matching.
SHERFACS is least culpable in this regard, since it retains hierarchical
relationships between conflict episodes, their phases, and actions within
those phases.
To best perform processual analysis, one would not proceed from numerical
data in the first place. Instead, one would proceed from the textual data
from which numerical events data are ordinarily derived. By expressing
events as textual propositions and representing them in a semantic network
formalism, 11 one can more easily capture contextual information as well as
information concerning the beliefs and intentions of actors.
Additionally, by constructing event models directly from textual sources,
one can construct and analyze comparatively alternative understandings of
events across political actors. Since they can also be expressed in prepositional
form, the normative and theoretical orientations of these actors can be

incorporated in proposition-based models (Duffy, 1988). Processual models
require temporal information, so one wants eventually to extract this in-
formation automatically from textual propositions. As discussed in a previous
section, this capability does not yet exist in RELATUS. Efforts to analyze
event processes need not await this development, however. One could simply
prespecify the temporal data when specifying the propositions.
Finally, with the generalization of the balance algorithm to the k-dimensional
case, additional dimensions may now be introduced. Timebase can easily
be extended to cover geographic as well as temporal coordinates. All that
is required are data on two additional dimensions-latitude and longitude.
Incorporation of geographical information with the temporal information
would provide us with a more refined tool for exploratory events data analysis.
We would have made time space a space-time space.
Notes
1. Richard H. Lathrop and John C. Mallery contributed to early conceptual
discussions at Gould/ AMI Semiconductors. Mallery also made valuable suggestions
for the TimebasejFeature Vector Editor interface. The examples reported here would
not have been possible without the Feature Vector Editor and the SHERFACS data
set. Frank L. Sherman provided access to the latter. Sherman, Lee Farris, Hayward
R. Alker, Jr., Kathleen Carley, and Renee Marlin coded the data. The author thanks
Symbolics, Inc., for providing the computational facilities on which Timebase was
developed.
2. A query might not be directly specified by the user. For example, in making
a logical inference, the system might query itself whether there is some x such
that x is a person and x did not want to recognize some y.
3. The axes are not shown in Figure 16.3.
4. Richard Lathrop and I worked out this reconceptualization of the temporal
domain in the summer 1985. Simultaneously, Jean-Fran~ois Rit developed a very
similar reconceptualization. This is not surprising, since we all stood on Alien's
shoulders. It should be noted, though, that Rit (1986) first brought the reconcep-
tualization to publication.
5. Finding the range consumes time proportional to log2N, while collecting all A
intervals consumes time proportional to A.
6. The row of intervals at the top of the temporal octant represents conflict
phases that have not yet terminated.
7. Presently, durations are specified on a separate unidimensional binary tree.
However, now that the balance procedure has been extended to the k-dimensional
case, a single three-dimensional can perform this task as effectively.
8. I am indebted to Dina Zinnes for pointing out the utility of Timebase for this
application.
9. Timebase can easily be made to be more finely grained in time. This would
have utility for other applications but doubtless not for international events analysis.
404 Gavan Duffy
10.Analyses of contagions require nearest-neighbor searches in the space. I am

indebted to Stuart Bremer for this observation. Procedures to support such searches
have been implemented and extended to the k-dimensional case.
11. Semantic networks are computer graphs that represent relationships between
propositions as well as relationships between the lexical items within propositions.
For a seminal work on semantic networks and their application in modeling textual
proposition, see Quillian (1968).
Bibliography
Adel'son-Vel'skii, G. M., and E. M. Landis. Dok/ady Akademia Nauk SSSR, 146
(1962): 263-266. English translation in Soviet Mathematics, 3 (1962): 1259-1263.
Alien, James F. "An Interval-Based Representation of Temporal Knowledge." /JCA/-
81: Proceedings of the Seventh International Joint Conference on Artificial
Intelligence. Vancouver, August 1981, pp. 221-226.
- - - . "Towards a General Theory of Action and Time." Artificial Intelligence. 23
(1984): 123-154.
Alien, James F., and Johannes A. Koomen, "Planning Using a Temporal World
Model," /JCA/-83: Proceedings of the Eighth International Joint Conference on
Artificial Intelligence, Karlsruhe Germany, August 1983, pp. 741-747.
Bennett, James P. "Data Stories: Learning about Learning from the U.S. Experience
in Vietnam," in Donald A. Sylvan and Steve Chan, eds., Foreign Policy Decision
Making (New York: Praeger, 1984).
Bentley, Jon L. "Multidimensional Binary Search Trees (]sed for Associative Searching."
Communications of the ACM. 18 (1975): 509-517.
Bentley, Jon L., and J. H. Friedman, "Data Structures for Range Searching." Computing
Surveys. 11 (1979): 397-409.
Bentley, Jon L., and H. A. Maurer, "Efficient Worst-Case Data Structures for Range
Searching." Acta lnformatica. 13 (1980): 155-168.
Comrie, Bernard. Aspect. Cambridge: Cambridge University Press, 1976.
___ . Tense. Cambridge: Cambridge University Press, 1985.
Duffy, Gavan. "Theory and Text: How Textual Analysis Can Help Build International
Relations Theory." Paper presented to the International Studies Association, St.
Louis, 1988.
Feierabend, I. K., R. L. Feierabend, and J. S. Chambers. "Transactional Data Bank
of International HostilityI Amity Events." Mimeograph. lnter-<Jniversity Consortium
for Political Research. Ann Arbor. June 1972.
Hornstein, Norbert. "Towards a Theory of Tense."Linguistic Inquiry. 8 (1977): 521-
557.
_ _ . "The Study of Meaning in Natural Language: Three Approaches to Tense,"
in Norbert Hornstein and David Lightfoot, eds., Explanation in Linguistics: The
Logical Problem of Language Acquisition (London: Longman, 1981), pp. 116-
151.
Mallery, John C., and Frank L. Sherman. "Beyond Correlation: Reconstructing The
Relational Structure from SherFACS Data." Paper presented to the International
Studies Association, St. Louis, 1988.
Moon, J. Donald. "The Logic of Political Inquiry: A Synthesis of Opposed Perspectives,"

in Fred Greenstein and Nelson Polsby, eds., Handbook of Political Science, vol.
1 (Reading, Massachusetts: Addison-Wesley, 1975), pp. 131-228.
Quillian, M. Ross, "Semantic Memory," in Marvin L. Minsky, ed., Semantic Information
Processing (Cambridge: MIT Press, 1968), pp. 227-270.
Reichenbach, Hans, Elements of Symbolic Logic (New York: Free Press, 1947).
Rit, Jean-Fran..ois, "Propagating Temporal Constraints for Scheduling." AAAI-86:
Proceedings of the American Association for Artificial Intelligence. Philadelphia,
August 1986, pp. 383-388.
Sherman, Frank L. "Quarrels and Disputes: An Expansion of the Phase/ Actor
Disaggregated Butterworth-Scranton Codebook." Mimeograph. Miami University,
Miami, Ohio, December 1985.
- - - . "SherFACS: A New, Cross-Paradigm, International Conflict Dataset." Paper
presented to the International Studies Association, St. Louis, 1988.
Vere, Steven A. "Planning in Time: Windows and Durations for Activities and Goals."
IEEE Transactions of Pattern Analysis and Machine Intelligence. 5 (1983): 246-
267.
---."Temporal Scope of Assertions and Window Cutoff." IJCA/-85: Proceedings
of the Ninth International Joint Conference on Artificial Intelligence. Los Angeles,
vol. 2, 1985, pp. 1055-1059.
Woisetschlaeger, Erich F. A Semantic Theory of the English Auxiliary System. Ph.D.
dissertation. Massachusetts Institute of Technology. Department of Linguistics
and Philosophy. 1976. Reproduced by the Indiana University Linguistics Club.
December 1976.
Yip, Kenneth Man-Kam. "Tense, Aspect, and the Cognitive Representation of Time."
/JCA/-85: Proceedings of the Ninth International Joint Conference on Artificial
Intelligence. Los Angeles, vol. 2, 1985, pp. 806-814.
ABOUT THE BOOK AND EDITOR
For well over a decade researchers in international relations have sought ways
to combine the rigor of quantitative techniques with the richness of qualitative data.
Many have discovered that artificial intelligence computer models allow them to do
just that. Computer programs modeling international interactions and foreign policy
decision making attempt to reflect such human characteristics as learning, memory,
and adaptation. In this volume of original essays, distinguished scholars present a
comprehensive overview of their research and reflect on the potential of artificial
intelligence as a tool for furthering our understanding of international affairs.
The contributors take a broad look at the early stirrings of interest in artificial
intelligence as a potentially useful method of political analysis, exploring such topics
as intentionality, time sense, and knowledge representation. The work also focuses
on the current state of artificial intelligence and examines its general areas of
emphasis: international interaction, decision making groups, and cognitive processes
in international politics. The contributors represent a cross section of different
approaches to using artificial intelligence and reflect the major research programs
across the country in this new international relations subfield.
Valerie M. Hudson is assistant professor of political science and director of

graduate studies at the David M. Kennedy Center for International and Area Studies
at Brigham Young University.
407
INDEX
AAAI. See American Association of Third World and, 6

Artificial Intelligence Al-type modeling, 202-203, 217(n6).
Accuracy, 372 See also Models
assessing, 211-213 AI winter, 10, 12-13
Act-concepts, 317 Alker, Hayward, 294
Act-descriptions, categories of, 313- work of, 1, 5, 6, 7(n4), 102-103
314 Alternative development strategies, 279
Action-descriptions, categories of, 313- Alternatives, packages of, 42
314 Ambiguity, 277
Actions, 319 intentional, 157
acts and, 314-317, 323 American Association of Artificial
description of, 312-313 Intelligence (AAAI), 12, 13, 122(n8)
PAL and, 230-231 Analogies, 20, 121, 171, 172, 343(n17),
patterns of, 324
349
reaction and, 315
adduced, 46
Actors
ANALOG system, 106
decision-making, 267
Analysis
external, 196
content, 106, 347, 371
motivation, 200-202
political, 245-246, 374 correlational, 14
problem-solving, 246, 254-255 counterfactual, 150, 208, 242-
rational, 129, 284, 294, 311 243(n 13), 264, 306
role of, 215 cross-aggregate national, 202
Acts discourse, 122(n7)
actions and, 314-317, 323 factor, 25
description of, 312-313 role, 146
social act schemata and, 324 sensitivity, 242-243(n13)
Adaptation, 19, 57, 276 syntactic, 348, 354-356
Al. See Artificial intelligence text, 98, 110, 115
AI/IR, 1, 16 Anderson, John, theory of, 294
data and, 25-26 And/ or graph, 63, 69
developments in, 3, 5, 9-10, 13-22 Archival materials, collecting, 233, 291,
models, 18-19 294, 326, 329-332
paradigms and, 26 Argument
problems with, 4, 10 connectives, 374
QIP and, 4 counterfactual, 302-303
rational choice and, 24-25 structure of, 16
409
410 Index
Artificial intelligence (AI), 40, 106, 187, Belief systems, 108, 113, 115, 118,
221, 241, 282-283 123(n10), 132, 145(n2), 353, 354,
adapting, 58-59, 283 356, 357
case-based reasoning and, 67 Bennett, James, 5, 7(n4)
concepts of, 283 Black box, 36, 240, 241, 284
description of, 5, 9-13, 59, 61 "Blocks worlds" program, 68
education and, 280-282 Books, AI, 12
information-processing research and, Boynton, G. R., 6
46-47 Brezhnev, Leonid, classification by, 347
path-dependent process and, 57 BOGGY, description of, 281
political science and, 23-26, 36, 59, Bureaucracy, machine replacement of,
82(n3), 282-283, 285-286, 307 189
rational science of, 23-25, 57-59 Butterworth, Robert, 99
temporal reasoning and, 149
See also AI/IR CAI. See Computer-assisted instruction
Artificial stupidity, 10, 22 Case-based models, 248
"As if" reasoning, 24, 36 description of, 267-268
flaw in, 311 rule-based systems and, 67-75
See also Reasoning See also Models
Axioms, 192(n17) Case-based reasoning (CBR), 57, 67,
patterns and, 170-172 83(n 12), 241, 279, 280
See also Knowledge, axiomatic description of, 71-73
development of, 56
Bach and Harnish Speech Acts, information-processing model and,
description of, 371 70-71
BCOW. See Behavioral Correlates of subtasks and, 74
War See also Reasoning
Banerjee, Sanjoy, 6 Case study approach, 20, 121(n4), 202
Behavior Categories, 127
decision-making, 246, 286 abducting, 329-332
decision on, 197-198 abstract, 363
international, 172 base, 363
nation-state, 194-195 CBR. See Case-based reasoning
patterns in, 2 Chronological minimization, 161-162
predicting, 188, 212-216 Chronology. See Sequences
political, 98, 128, 169-174 Classical planning framework, 67
properties, 212-213 Classification
realism and, 284-286 concept reference, 363
role, 195 description of, 347-348
rules and, 204, 205, 285 hierarchical, 365-366
See also Foreign policy behavior instance, 363, 365
Behavioral Correlates of War (BCOW), test, 367
23, 170, 174, 176-179, 183, 186, See also Lexical classification
190(nn 8, 9) Classification modules, description of,
Behavioral output, replication of, 239 256-257, 259-260, 266
Belief-system examiner, 361, 363, 366- Coding, 116-117, 217(n10), 343(n8)
368, 376(n4) schemes, 45-46
Index 411
Cognition Condition-action pairs. See If . . . then

bounded rationality of, 306 statements
computational models and, 188 Condition-event patterns, sample of,
research on, 37-38, 294-296 173
theories of, 296, 307 Conditions
Cognitive action, model of, 296 assessing, 238
Cognitive conceit, 38 definition of, 173
pervasiveness of, 48-49(n4) Conflict and Peace Data Bank
Cognitive limitations, description of, (COPDAB), 179, 209, 371
40-43 Constraint-interpreting reference
Cognitive psychology, 14, 15, 36, 40 system, 354, 378(n22)
Cognitive science, 58 Constructs
convergent and divergent, 372
emphasis on, 36
Context sensitivity, 117, 150
Commitments
Contingency, sanction and, 138-142
creating, 141
COPDAB. See Conflict and Peace Data
predicting, 215
Bank
Comparative Research on the Events of
Coreferential capabilities, 117
Nations project (CREON), 23
Creativity
Complementarity, 24 computers and, 19
Compliance conditions, 150 politics and, 127
Comprehension, duplicating, 188 CREON. See Comparative Research on
Computational models, 4, 7(n1), 19, 37, the Events of Nations project
40, 128-134, 163, 169, 170, 222 Cross-fertilization, benefits of, 282
advantages of, 149-152 CSRL. See Conceptual Structures
cognitive ability and, 188 Representation Language
description of, 10-11, 35-36 Cues
information-processing insight and, external, 38, 45-47
42-43 internal, 38
ontological differences and, 152-153
problems of, 44-48, 48(n1)
DARPA. See Defense Advanced
temporal reasoning and, 162
Research Projects Agency
using, 221, 239, 327
Data, 25-26
See also Models
constructing, 329, 332-335
Computational side effects, 140 eliciting, 61
Computer-assisted instruction (CAI), events, 189-190
description of, 280-282 patterns in, 60
Computers, creativity and, 19 testing, 179
Concepts, 266 verbal, 49(n9)
deciding on, 283 Decision making, 49(n14), 246, 277,
learning, 76, 77-81 286, 318-319
modification of, 81 context-dependent, 226
presentation of, 260, 262 organizational, 14
stock of, 81 political development and, 274-275
Conceptual Structures Representation steps in, 225
Language (CSRL), 262 structure of, 40-42
412 Index
Decomposition, 146(n7) EBL. See Explanation-based learning,

levels of, 142 description of
Deeds, power of, 98 Education, 280-282
Defense Advanced Research Projects models for, 286
Agency (DARPA), 12, 27(n3) Efficacy conditions, definition of, 131
Deictic context, 356, 377(n20) Efficiency, context-sensitive, 117
Deliberative reference processing, Eidetic representation, 350, 352
description of, 352 Embarrassment, escaping, 155
Delta rule, 178, 187 EMYCIN, 61
DENDRAL, 61, 64 Energy-related decision making, 245,
Descriptive behavioral science, 58 249-251, 257, 259
Design Specialists Planning Language foreign policy and, 251, 254
(DSPL), 262 JESSE and, 251, 253, 264, 266, 269
Deterrence, 131-133, 138-139, 146(n9) task of, 254-255
audience effects of, 142 Environment, 132
computational modeling of, 128-134 anarchic, 284
definition of, 129, 140 Epistemology, evolutionary
constructivist, 101, 118
extended, 133-134, 143-144
Event-content, implications of, 152
internal structure of, 129-130
Event creators, 128
mutual, 142-143
Events, 128, 402
Developing countries
chronology of, 299
decision making in, 273-278
coding of, 217(n10)
models for, 278-280, 286
deletion of, 154-155, 186
Differential perception, modeling, 374 description of, 153, 173
Direction of Trade (IBRD and IMF),
insertion of, 154-155
209 novel, 127, 173
Disambiguation, 355, 369, 371, unreported, 236
376(n3), 378(n28) See also Nonevents, accounting for
Discourse, 122(n7) Event sequences, 176-179, 187
structure of, 16 description of, 153-154
Distributed information processing, importance of, 189
262-263 See also Sequences
Dissuasion, 129 Example-counterexample protocol, 175,
computational modeling of, 134-136 187
Documents. See Archival materials, Expert systems, 16
collecting design of, 66
Domain using, 11
choosing, 249-250 Explanation-based learning (EBL),
independence, 117, 350 description of, 56, 75-82
nature of, 268 Explanatory schema acquisition, 77-78
DSPL See Design Specialists Planning Explicit language processing,
Language description of, 352
Dual track information processors, 194 Explicit texts, 115, 123(nn 10, 11)
Duffy, Gavan, 6, 7(n4) sample of, 104-105
Dynamic knowledge base, 56 Expressibility, problems with, 1-2
Index 413
Extensibility, 117 Hypotheticals, using, 279-280

concepts of, 144
External environment, features of, 196- ICPSR. See Inter-university Consortium
198 for Political and Social Research
If . . . then statements, 15, 17, 20, 60,
159, 236
Feature Vector Editor, 387, 394, 402
IMF. See International Monetary Fund
Feigenbaum, Edward, 61
Immediate reference processing,
Fifth Generation Project, 12
description of, 352
Fisher, Waiter, 294
Incremental chooser, 43
Foreign policy behavior, 209, 216(n3),
Indexical categories, 151, 261-262, 266
260-263
using, 255, 257
differentiating, 213, 217(n5)
Inferences, 158, 260
measurable properties of, 201-202
analogicalfprecedential, 115
predicting, 215
contingent, 159
See also Behavior
drawing, 65
Fox, John, 285
nonmonotonic, 159
Framing, 39, 44, 202, 285, 347, 367
practical, 386
Friedman, Milton, 24
Information
cues, 44
Generalization, 265, 269, 349, 363, linking, 188
368, 373, 403 pattern recognition and, 172-174
General Problem Solver, description of, Information processing, 46-48, 58,
368, 371 247-248
Glass box approach, description of, computational models and, 42-43
281-282 description of, 245-246, 254
GNOSCERE, 353, 354, 357, 376(n4) distribution of, 263
Grounded strategy, 6, 331, 333, limits on, 43
343(n13) research on, 37-47
Group information processing, 249 Information-processing model
case-based reasoning and, 70-71
Haas, Ernst B., 99 political content in, 247
Have!, Vaclav, 97-98, 108, 116 See also Models
Hermeneutics, 99, 350-352, 376(n3) Initialization, frequency-based, 182
computational, 6, 119 Inspection, description of, 348
development of, 118 Instability, 223-225, 241 (n4)
Heuristics, 15, 80, 131, 373-374 Intelligibility, 144, 146(n9)
information-processing research and, description of, 127-128
47-48 lntentional-inferencing, 150
temporal inferencing and, 163 Intentional relations matrices, 322-323
using, 38, 41, 43-45, 75-76, 202, International Joint Conference on
279-280, 285 Artificial Intelligence, 13
Hierarchical expression, 389-390 International .Monetary Fund (IMF),
Hilbert, David, 11 209, 274
Historical structures, 317-318 International Political Science
reproduction of, 323-324 Association, 13
Hurwitz, Roger, 371 Interpersonal relations, 77-78
414 Index
Interpretations, 127, 144 deductive, 171

Interpretive prejudices, adjusting for, engineering, 70
115 experiential, 268-269
Interpretive triples, 47 level, 296
as counterfactual arguments, 302- modification of, 18
303 prior, 49(n4)
description of, 301-302 representation, 11, 15-16, 353
flexibility of, 306-307 structure, 67, 176
as predictions, 303-305
lntersentential reference, process of, Language
355 artificial, 9
Inter-university Consortium for Political dynamic, 118, 131
and Social Research (ICPSR), 26 gap, 282-284
Interventions, analyzing, 20, 81, 99-101 generic task, 248
Intervention type situations, 206-208 information-processing, 246, 248
Intuitive attributions, diversity of, 45 theory and, 282
Intuitive problem solving, using, 38-40 See also Natural language
Language-based processes, 81
Lasswell Value Dictionary, 378(n28)
Japanese Energy Supply Security
description of, 368
Expert (JESSE), 14, 18, 189, 264
LDM. See Learning-driven models
architecture of, 255-257, 259-260
Learning, 27(n9), 59, 75-82, 307
case-based reasoning and, 267
AI/IR models and, 18-19
development of, 6, 66, 245, 265-266
analytic, 75
function of, 247-248, 251, 253-254,
empirical, 75
269-270
by experience, 285
inference mechanisms of, 260
experimental, 57
rule-based system and, 267
organizational, 19
substantive level of, 250
political, 80-82
validation of, 248, 263-265, 269
See also Machine learning
JESSE. See Japanese Energy Supply
Learning-driven models (LDM), 28(n 13)
Security Expert
description of, 21-22
Jervis, Robert, 143
See also Models
Job, Brian, 6
Levenshtein distance, 170, 181, 182,
Johnson, Doug, 6
185
"Judgment, Policy and the Harmony
computing, 179-180
Machine" (Fox), 285
event substitution and, 186
Kahn, Herman, 141 weakness of, 184
Kahneman, Daniel, 14-15 Lexical classification, 349, 368-370,
Keesing's Contemporary Archives, 209 374-375
Keohane, Robert, 170, 310 description of, 363, 365-368, 371
Knowledge hierarchical, 365-366
acquisition, 192(n 17), 279 See also Classification
axiomatic, 171-172 Lexical disambiguator. See
commonsense, 122(n7), 195, 349 Disambiguation
declarative, 295, 307 Limits to Growth (Meadows), 276
Index 415
Linguistic encoding, 130 unlimited, 14

Linguistic/ interpretive modeling working, 40-41, 45, 171
philosophy, description of, 116- See also Recall
119 Memory research, 40, 83(n13)
Linguistics, computational, 99, 123(n 12) Mental biases, 43-45
LISP, 11, 13, 27(n9), 106, 117, Metarules, 227, 230
123(n 12), 132, 262, 354, 355, 356, Methodologies, 202-203
365, 367, 368, 376(n4), 377(n12) restrictions of, 58
Literal language processing, 121(n6) understanding, 283
description of, 352 Modal logic
Literal texts, 106-107, 115, 123(nn 10, description of, 159-160
11) monotonic, 160-161
sample of, 104-105 nonmonotonic, 159-163
Logic temporal reasoning and, 163
direct-action response cybernetic, See also Logic
240 Model of muddling, 43
nonmonotonic, 159 Models
precedent, 48(n3), 121(n4), 373 arms race, 14
reflective, 121(n4) building, 37, 202-203, 276-280,
standard, 159 330-331
See also Modal logic case-specific, 66
Logical form, 111, 356, 359 cognitively plausible, 44, 47, 48(n1)
LOOPS, 262 computational, 7(n1), 36, 37, 128-
Luttwak, Edward, work of, 57, 59-60, 134, 149-152, 163, 169, 222, 327
62, 64, 69-71, 74-75
constitutive, 328, 330-331
dynamic, 14, 18-22, 118, 131
Machine intelligence, 187-189
game theoretic, 14
Machine learning, 187, 190(n6),
grounded computational, 335-336
191(n17)
ideal, 22
crises and, 182-184
explanation-based learning and, 77- inaccessibility of, 282
80 linguistic, 25
pattern recognition and, 174-176 modified cybernetic, 222, 225-226,
See also Learning 240
Macro-operators, 76, 77 processual, 328, 402, 403
MACSVMA, 64 production system, 16, 17
Mallery, John, 6, 387 proposition-based, 403
Mankind at the Turning Point reproduction, 312, 323-324
(Mesarovic and Pestel), 276 single-actor, 225
Maximizers, 152 static, 14
Maxims, 75-76, 207 testing, 209-210
Mefford, Dwain, 5, 294 See also Text modeling
Memory Modification, 21-22, 81
long-term, 71, 171-172, 174 Multiple indices, 261-262, 268
research on, 40, 83(n 13) Multisense processing, description of,
short-term, 40-41, 43, 71, 190(n2) 352
structures of, 72 Mutual deterrence, 142-143
416 Index
MYCIN, 61, 64, 65 Object orientation, 283

Object relation, 357
Olafson, Frederick, 294
Narrative construction, 103, 106,
Ontological assumptions, 348
121(n4), 292-294, 297-298, 306-
importance of, 151-152
307
Operational code, 27(n5)
theory of, 296
constraints of, 17
Narrative data, sample of, 100-101
using, 41
Narratives, 106, 308(n3)
Operators, 76
conflict, 358, 360
truth-functional, 159
reconstructed, 46
OPS5, 61, 82(n8)
research on, 294-296
Optimization, 24
National interests, 307
Options, limited, 42
National role conceptions, defining,
Organizational performance, 211
195-196
analysis of, 212
National security, economic, 250
Organized complexity, 3
National Security and U.S. Policy Organized simplicity, 3
Toward Latin America (Schoultz), Ortony Affect Lexicon, description of,
223 371
National Security Council (NSC) 30 Outcomes, 173, 280, 282
(document), 321-322 negative, 44
National Security Decision Directive validity, 25-26, 210, 264-265
(NSDD), 138
Natural language, 348, 353, 360, PAL. See Political action language
376(n11). See also Language Paradigms, proliferation of, 26
Natural language processing (NLP), 9, Parsing, 116-117, 121(n5), 122(n8),
16, 23, 59, 102 123(n11), 309(n23), 347, 353-355,
models of, 352-353 361, 368
RELATUS and, 110, 120(n3) description of, 110-111, 113, 115
study of, 5 Partially ordered event structures
Neural networks, 27(n6), 178, 187, (POES), description of, 184-186
190(n7) Path-dependent process, 19
Newell, Alien, 14, 60 description of, 57
Neweii-Simon Information Processing political learning and, 80-82
Model, 15, 72 Pattern recognition, 46, 174-176
NLP. See Natural language processing axiomatic reasoning and, 172
Nonevents, accounting for, 229. See description of, 75, 169
also Events long-term memory and, 171-172,
Nonwar crises, 180-184, 186 174
NSC. See National Security Council political behavior and, 170-174
NSDD. See National Security Decision problem solving and, 169-170
Directive Perceptual consensus, 314-315
Nuclear deterrence. See Deterrence Perestroika, impact of, 97, 115
Nuclear Weapons Employment Policy Personal computer, impact of, 11-12
(NUWEP82), 138 Phase-line combination, 335-336
NUWEP82. See Nuclear Weapons amalgamated, 341
Employment Policy description of, 333-334
Index 417
Phasing, description of, 333-334 Posture shifts, 231-233, 235-236,

Phenomenological representation, 240-241
description of, 350 Power drama
Phillips, Warren, 6 roles in, 198-200
Piaget, Jean, 319 script of, 215-216
Planning, 74, 78, 83(n16) Precedents, 373
activity of, 67 concept of, 20-21
automatic, 150 Predictability, 169, 213-215, 311, 373
hierarchy, 252 assessing, 210
long-range, 275 Predictions, interpretive triples as, 303-
maintenance, 266 304
reactive, 255, 265 Prejudicial horizons, 98
redefining, 67-68 Primitives
routine, 254-255 conceptual dependency, 351
Plausibility, 222, 239, 261, 263, 269, relations, 230
297-299, 333 semantic, 230
Plausible connections, establishing, Prior affect, 206-209
300-301, 304-305 description of, 200-201
Plausible deniability, 154 Prisoner's dilemma model, 311
POES. See Partially ordered event description of, 360, 371
structures, description of See also Models
Pointer propagation problem, 388-390 Problems
Policy domain, 278-279 framing, 15, 285
Policy identifying, 38-39
anticipatory, 250-251 subject of, 198-199
domain, 278-279 Problem solving
line, 332-334 architecture of, 71
Political action language (PAL), 234, assumptions about, 14
235, 241-242(n7), 242(n8) pattern recognition and, 169-170
actions and, 230-231 research on, 48(n2)
UNCLESAM and, 229 Problem-solving capabilities, 247, 277
using, 227 Production systems, 16, 17, 82(n4)
Political development description of, 60-61
decision making and, 274-275 See also Rule-based systems
history of, 278 Programming, AI style of, 11-12
problem-solving capabilities and, 277 Prolog, 11, 82(n8), 83(n17), 123(n12),
Political language, 246 325(n14)
acquiring, 81 Prepositional attitudes, 150
Political messages, sending, 152 Propositions
Political present, 155 conceptual, 386
Political science, AI and, 23-26, 59, contextual, 386
82(n3), 282-283, 285-286 description of, 295-296
Political structures, 245, 276 Prospect theory, 15
Politics Among Nations (Morgenthau), Psycholinguistics, 351
284 Public plans, notion of, 68
Pool, lthiel, 106 Purkitt, Helen, 5
418 Index
QIP. See Quantitative international data for, 344(n20)

politics differentiating, 334
Quantitative international politics (QIP), identifying, 331
1, 3, 4 negative, 342(n3)
Reeves, Richard, 127
Randomness, 169, 185 Reference specifications (ref-specs),
problems with, 2-3 102, 356, 357, 359, 370, 377(n18),
Rational choice (RC), 23-25, 28(nn 13, 389-390
14), 36, 311-312 description of, 354
Rationality, 194, 278, 310-312 graphs, 113
assuming, 284-285 Referential integration, 123(n 11)
political, 349 description of, 348
procedural, 349 Referential opacity, 351
RBS. See Rule-based systems Reflective perception, 323
RC. See Rational choice
Reftectivist approach, description of,
Reaction, action and, 315
310
Read, Stephen, 294
Regime orientation, concept of, 196
Reagan, Ronald, 291
Relational network, 199-201, 209
classification by, 347
Relative capabilities, 209
Realism, impact of, 282, 284-286
Reasoning, 57, 171, 187
RELATUS, 48(n3), 49(n9), 100, 108,
analogical, 19-20, 73, 106, 347
114, 120(n1), 351, 355-358, 361,
compiled, 18, 21-22
367, 376(nn 9, 10), 377(nn 12,
deductive, 19
21), 387, 402, 403
by example, 73
contribution of, 99, 119
knowledge-based, 260-263
logical, 14, 188 coreferential capabilities of, 117
memory-based, 11 description of, 6, 98-99, 101-102,
models for, 1 106, 120(n2), 352-353
nonmonotonic, 158-159 interpretive prejudices and, 115
political, 14-16 limitations of, 121-122(n6)
reconstructive, 158 methodological capacities of, 121(n5)
richness of, 2 natural language processing and,
script-based, 241 110, 120(n3), 347, 376(n4)
temporal, 149, 162, 163 parsing with, 110-111, 113, 115
See also "As if" reasoning; Case- referential integration and, 123(n 11)
based reasoning text modeling by, 101-102, 113,
Recall, 171 116-119, 349-350
associative, 188, 189, 190(n3), RELATUS Natural Language System,
192(n17) 347, 376(n4)
See also Memory Reliability, 371-372
Reciprocal causal relationships, 116 Replication
Recommendations, 330, 332, 333, 336, capacity for, 242(n13)
342(n6), 343(nn 12, 17) importance of, 237-239
categorized, 338-341 Reprisal, definition of, 129
computational model of, 327-328, Resource constraints, importance of,
335 250
Index 419
Retrieval, analogicalfprecedential, 115 limitations of, 70

Risk orientation, importance of, 194 modification of, 21-22
Robinson's Resolution Principle, using, stability assessment, 235
68 time-dependent, 227, 229, 236
Roles, 199-200 types of, 227
coding of, 217(n10) use of, 18, 235
concept of, 201 Rummel, Rudolph, 2
identification of, 197
source, 199, 206 Salience, 207-209, 225, 233, 276
subject, 199, 206 description of, 200-201
Role set, 195 Sanctions, 146(n8)
Role theory, 198 contingency and, 138-142
description of, 195 Satisflcers, 152
Root-Bernstein, Robert, 3-4 Schank, Roger, 72
Routine hypothesis-testing strategy, Schema, 11, 41, 247, 295, 296, 323
333. See also Tests acquiring, 77-78, 80
Rule-based decision models, 235, 248, action, 318-320
327 reinforcement of, 322
description of, 266-267 scripts and, 322, 324
using, 239-241 See also Social act schemata
Schemes
See also Models
learning, 18
Rule-based systems (RBS), 13, 27(n6),
positivist, 312
189, 203, 210, 267, 348
SCHOLAR, description of, 281
advantages of, 208-209
Schoultz, Lars, 223, 224
alternatives to, 56
Schrodt, Philip, 1, 5, 294
case-based systems and, 67-75
Scientific desiderata, 3
description of, 16-18, 59-67, 221
Scranton, Margaret, 99
first-generation, 60-63, 73
Scripts, 11, 78, 83(n16), 195, 283,
limitations of, 56, 70, 240-241
295, 317-319
replicability and, 237-239
acquiring, 77
second-generation, 64-65
calculation of, 216
third-generation, 65-67 concept of, 72
See also Production systems description of, 312
Rules, 20, 62, 283 schemata and, 322-324
acquiring, 76 Security, 268, 321-322
assessment, 227, 229 Selective dimensions, nature of, 197
behavior and, 285 Self-modification, 21, 24, 27(n9)
content-dependent, 240 Self-selection process, 249
decision, 127, 215, 227, 230, 236, ~mantle content analysis, 350
285 construct validity and, 372
informal, 17 description of, 347-348, 371-372
interaction, 204, 205, 207 Semantic inversion, 360-361, 368
interpretive triples as, 307 Semantic networks, 106, 404(n11)
isolation, 204-208 Semantics, 152, 356
justifying, 67 lexical-interpretive, 101, 118, 350-
language of production, 266-267 352
420 Index
Semantic texts, 106-107 Social act schemata

Senate Foreign Relations Committee, acts and, 324
298, 305, 307 description of, 319-320
hearings of, 291-292, 295-297 reinforcing, 320-322
narrative construction and, 292-294 scripts and, 323
Sentences See also Schema
generating, 377-378(n22) Social clock, 157
parsing, 110-111, 113, 115 Social perception, process of, 312-317
Sentential constraint poster, description SOPHIE, description of, 281
of, 355 Spatial images, 190(n1)
Sequences, 170, 386 description of, 295
constructing, 298-299 Speech acts, 136-143, 364-365, 368-
current, 171 369
examining, 298-299 analysis of, 145(n4)
historical, 171 application of, 145(n5)
ideal, 186 lexical classification of, 365, 370
temporal, 154, 264 retractive, 367, 370
See also Event sequences Stability, 242(n10), 251, 324
SHERFACS, 99, 103, 121(n5), 360, assessment of, 235, 236
387, 394, 396, 397, 402 classificational, 371-372
SHERFACS Actions, 371 Standard antinomies, 333
SHERFACS International Conflict Standard cybernetic decision-maker,
Management Dataset, 371 222
Sherman, Frank, 99, 102-103, 387 Static knowledge base, 56
Short-term reactions, 275 Statistical Yearbooks (UN), 210
SHRDLU, 68 Stereotypes, 223-224
Simon, Herbert, 14 Stored plans
Simplicity, 224 retrieval of, 255
Sincerity conditions, 134-135, 140 selection of, 262
Single sense processing, description of, See also Planning
352 Story grammar concepts, 186
SlOP, 67, 138 Strategic Computing Initiative, 12
Situational predisposition model, 203, Strategic environments
205, 209, 210 time horizons and, 156
accuracy of, 212-213 uncertainty of, 158-162
description of, 195 Strategic manipulation, temporal
See also Models distinctions and, 155
Situational Predisposition Strings, constructing, 296-300
Operationalization Manual, 209 STRIPS project, 76
Situations development of, 68
categorizing, 197, 328 Strong simulations, 149
description of, 150, 328 Strong theory, 69-70
role of, 199 striving for, 4-5
Situation type, concept of, 201 Structural adjustment programs (IMF),
Slovic, Paul, 14-15 274
SOAR, 15, 19 Structural historiography, 324
Index 421
Structural realism, impact of, 284-286 problems of, 329

Structure, 21-22 Theoretical relevance, 239
Subject relation, 357 Theories
Subjects, reproduction of, 322-323 evaluation of, 284
Sylvan, David, 6, 7(n4) language of, 282
Sylvan, Donald, 1, 6 Theory of generic tasks, description of,
Synchronization, description of, 157- 253-254
158 Theory of International Politics, The
Syntax (Waltz), 284
interface, 355-356 Third World, AI/IR and, 6
temporal representation and, 153 Thorson, Stuart, 1, 5, 7(n4)
Time
Tamashiro, Howard, 5 computational interpretations of, 163
Taylor, Charles, 4 diplomatic, 156
Technical conditions, definition of, 131 modeling, 162-163
Temporal contexts ordinal relationships of, 388
analyzing, 387 perception of, 153
description of, 295-296 political choices and, 386-387
Temporal indexation, motive for, 387- reasoning and, 388
388 reconceptualizing, 389, 392-393,
Temporal inferencing, 150 403(n4)
heuristics and, 163 sense, 158, 163
Temporal maneuverings, model of, 150 strategic, 158, 160
Temporal strategies, 153-158, 160 subjective, 150-153
Tests Timebase, 388, 398, 403(n9)
constructing, 209-210 description of, 387
split sample, 176 using, 394, 397, 402-403
Text modeling, 98, 99, 113, 117, Time-based activities, types of, 150
121(n4), 123(n12), 347, 358, 366 Time conceptions, content of, 152-153
analyzing, 108, 110-111, 113, 115- Time duration, 156-157
116, 350, 354 Time horizons
belief-system-specific, 102-103 description of, 155-156
constructing, 108, 110-111, 113, expanding, 162
115-116, 350, 353-354 Time space, 390, 391, 393
editor mode for, 361 Timing, definition of, 154-155
formula for, 102 Transducers, building, 143-144
future of, 119 Truth maintenance functions, 230
lexical classification for, 374-375 Truth-value, 159, 163
relevance of, 348-349 Tversky, Amos, 14-15
sample of, 100-101
See also Models UNCLESAM, 16-17, 241(n6)
Text processing, description of, 101- description of, 6, 221-222, 226-229,
102, 356-358 239-240
Text representation, 348-350 PAL and, 229-230, 241-242(n7)
Textual propositions, 402 replicability and, 238, 242(n 13)
Theoretical categories, 332, 335 using, 222, 229-240, 243(n 13)
abduced, 330-331 weaknesses of, 240-241
422 Index
Unilateral adjustments, 251 War crises, 186

Unilateral relationships, 223, 225 discriminating, 180-184
United Nations, 99-100 War Powers Act, 292, 303, 306,
United States Arms Control and 309(n26)
Disarmament Agency (USACDA), Weak simulations, 149
209 Weak theory, 69
Unorganized complexity, 3 Weight matrix, 178-180
USACDA. See United States Arms Weinberg, Gerald, 2-3
Control and Disarmament Agency WEIS. See World Event Interaction
Utility, assessing, 211 Survey
WEST, description of, 281
Winston, Patrick, 106
Validity, 248, 269, 306, 371 Word force, 115
conceptual, 210 Words
construct, 372 changed meanings of, 108
external, 222 impact of, 97-98
hypothesis, 372-373 World Dynamics (Forrester), 276
predictive, 373 World Event Interaction Survey (WEIS),
process, 25-26, 175-176, 187, 179, 186
190(n6), 217(n12), 264-265 World Military Expenditures and Arms
Variables Transfers (USACDA), 210
dependent, 201, 202
explanatory, 202 Zhdanov, Andrei, 313, 322

Valerie M. Hudson - Artificial Intelligence and International Politics-Westview Press (1991)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Valerie M. Hudson - Artificial Intelligence and International Politics-Westview Press (1991)

Uploaded by

Copyright:

Available Formats

Artificial Intelligence

and International Politics

Published 2018 by Routledge

Routledge is an imprint of the Taylor & Francis Group, an informa business

Copyright © 1991 Taylor & Francis

Library of Congress Cataloging-in-Publication Data

ISBN 13: 978-0-367-00371-5 (hbk)

About the Contributors vii

Introduction, Valerie M. Hudson 1

1 Artificial Intelligence and International Relations:

2 Artificial Intelligence and Intuitive Foreign Policy

3 Steps Toward Artificial Intelligence: Rule-Based, Case-Based,

4 Text Modeling for International Politics:

5 Reasoning and Intelligibility, James P. Bennett and

6 The Computational Modeling of Strategic Time,

7 Pattern Recognition of International Event Sequences:

8 Scripting International Power Dramas: A Model

9 UNCLESAM: The Application of a Rule-Based Model

10 Modeling Foreign Policy Decision Making as

11 Decision Making and Development: A "Glass Box"

12 The Expertise of the Senate Foreign Relations Committee,

13 Reproduction of Perception and Decision in the

14 Theoretical Categories and Data Construction in

15 Semantic Content Analysis: A New Methodology for the

16 Time Space: Representing Historical Time for

About the Book and Editor 407

Hayward R. Alker, Jr., is a Professor of Political Science at MIT.

Jennifer L. Milliken is a doctoral candidate in Political S.cience at the

The purpose of this volume is to provide a general yet relatively com-

Figure 1: Weinberg 's Classification of Subjects (From Weinberg, 1975, p. 18)

Ill. Orgaruud Complexity

studying international relations. This problem had numerous facets: inadequate

In Region I, which Weinberg calls the realm of "organized simplicity," or

It is not science's ability to reach solutions that needs to be accounted for,

recognition techniques to tease out similarities and dissimilarities in sequences

of the financial support provided by the Department of Political Science at

Foundation Conference on Dynamic Models of International Conflict, Boulder,

Artificial intelligence techniques have been used to model international

"artificial intelligence," including the widespread use of LISP and Prolog as

AI in the Early 1980s

Research in AI has always proceeded in parallel, rather than serially, with

of programmers and programming environments able to construct compli-

the popular press as attention turned away from AI to global warming,

AI/IR Research: A Framework

it is decidedly uncharacteristic of formal approaches to the study of

Patterns of Political Reasoning

acteristics predicted by utility maximization and statistical decision theories

of statistical techniques and the RC adoption of economic techniques: In

Static Modeling: Rule-Based Systems

IF U.S. Posture to the Dominican Republic government > 4

Stability Level > = 5

In short, rule-based political behavior is not an "as if" proposition: It

Dynamic Modeling: Adaptation and Learning

SOAR project. The content of an organization's compiled reasoning in turn

Consider, for example, the situation of a failure of reasoning by analogy

The Upshot: The Ideal Model

• The model would be consistent with the psychological literature on

AI and Contemporary Political Science

AI and Rational Choice

using preferences and constraints would be using a sledgehammer to kill

bottlenecks imposed by the available techniques. Common mathematical

AI/IR and Data

outcome validity would reproduce or predict behavior but make no pre-

E(A) = p(100) + (1-p)(-50)