Professional Documents
Culture Documents
edited by
Valerie M. Hudson
First published 1991 by Westview Press, Inc.
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.
PART ONE
CONCEPTUAL ISSUES AND PRACTICAL CONCERNS
PART TWO
AI/IR RESEARCH: INTERNATIONAL EVENTS AND
FOREIGN POLICY DECISION MAKING
V
vi Contents
PART THREE
AI/IR RESEARCH: THE DISCOURSE OF FOREIGN POLICY
vii
viii Contributors
1
2 Valerie M. Hudson
analytical treatment
COMPLEXITY statistical treatment F''>'>>'
',',. .,. .,. .,. .,. .,l
[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
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
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
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.
Artificial Intelligence
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
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.
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
AI and International Relations: An Overview 13
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.
18 Philip A. Schrodt
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.
22 Philip A. Schrodt
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.
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 International Relations: An Overview 23
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. . . .
AI and International Relations: An Overview 27
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
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.
AI and International Relations: An Overview 29
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
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
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.
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
AI and Intuitive Foreign Policy Decision-Makers 41
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,
AI and Intuitive Foreign Policy Decision-Makers 45
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).
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
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
AI and Intuitive Foreign Policy Decision-Makers 49
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.
AI and Intuitive Foreign Policy Decision-Makers 51
___ . 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.
Hillsdale, NJ: Lawrence Erlbaum.
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.
AI and Intuitive Foreign Policy Decision-Makers 53
___ . 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
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.
Hillsdale, NJ: Lawrence Erlbaum.
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
56
Rule-Based, Case-Based, and Explanation-Based Models 57
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.
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
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)
Figure 3.1: AND/OR Graph of Tasks Implicit in the Concept of Coup d'Etat
I I
I
by exploiting by compromising
geographical communications
dispersement AND
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.
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
Rule-Based, Case-Based, and Explanation-Based Models 67
(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).
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.
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
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
Rule-Based, Case-Based, and Explanation-Based Models 73
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
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
Rule-Based, Case-Based, and Explanation-Based Models 75
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.
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.
Rule-Based, Case-Based, and Explanation-Based Models 79
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)
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)
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)
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
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-
Rule-Based, Case-Based, and Explanation-Based Models 83
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
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
___ . 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
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.
Rule-Based, Case-Based, and Explanation-Based Models 93
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.
Rule-Based, Case-Based, and Explanation-Based Models 95
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,
CA: Morgan Kaufmann.
___ . 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
Kaufmann, pp. 245-250.
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
Kaufmann, pp. 451-458.
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
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
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.
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
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.
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
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
( BE-87) (nRTYRS-RRKOSI-1)
I'( BE-82)
~ (BE-88) (GER0-1)
.....
~
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.
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
A Tourist's Guide to RELATUS 113
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
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.
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
120 H. R. Alker, Jr., G. Duffy, R. Hurwitz, and J. C. Mallery
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-
A Tourist's Guide to RELATUS 123
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.
124 H. R. Alker, Jr., G. Duffy, R. Hurwitz, and J. C. Mallery
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.
126 H. R. Alker, Jr., G. Duffy, R. Hurwitz, and J. C. Mallery
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
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
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:
(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
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.
Reasoning and Intelligibility 135
"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
138 James P. Bennett and Stuart J. Thorson
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
Reasoning and Intelligibility 139
deterred believes that the would-be deterrer has prepared sanctions with
which to respond should the contingency (8) be fulfilled:
where C' will typically be similar to but not identical with (secret) C and
(9):
(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):
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:
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
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)
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.
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).
Reasoning and Intelligibility 145
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).
Reasoning and Intelligibility 147
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.
148 James P. Bennett and Stuart J. Thorson
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:
149
150 Howard Tamashiro
3. Since time strategies are our main interest here, we want time
conceptions that allow for the study of interesting time maneuverings.
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
154 Howard Tamashiro
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
156 Howard Tamashiro
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
Computational Modeling of Strategic Time 157
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
158 Howard Tamashiro
and raise the ante without risking the dangers of an irreversible "take-it-
or-leave-it" ultimatum.
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
Computational Modeling of Strategic Time 163
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.
164 Howard Tamashiro
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.
Computational Modeling of Strategic Time 165
AI/IR Research:
International Events and
Foreign Policy Decision Making
7
Pattern Recognition of International Event
Sequences: A Machine Learning Approach
Philip A. Schrodt
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
170 Philip A. Schrodt
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. Sequences of events;
2. The conditions at the time those events occurred.
Pattern Recognition of International Event Sequences 173
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).
Pattern Recognition of International Event Sequences 175
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
Pattern Recognition of International Event Sequences 177
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
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
Pattern Recognition of International Event Sequences 179
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.
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."
180 Philip A. Schrodt
Training Set
Training Set
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:
182 Philip A. Schrodt
Minimum Weights
Maximum Weights
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
184 Philip A. Schrodt
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
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]
[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.
186 Philip A. Schrodt
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.
188 Philip A. Schrodt
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
Pattern Recognition of International Event Sequences 189
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.
Pattern Recognition of International Event Sequences 191
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:
192 Philip A. Schrodt
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
Pattern Recognition of International Event Sequences 193
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.
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.
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
Diagram: Actor Source Subject
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
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
Diagram: Actor Source Subject
A X X
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
202 Valerie M. Hudson
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.
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
Scripting International Power Dramas 207
~
~
(lD48)
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,
210 Va/erie M. Hudson
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-
Scripting International Power Dramas 211
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
212 Va/erie M. Hudson
(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
Table 8.4: Number of Distributional Peaks of Frequent Behavior and Percent of Those
Behaviors Predicted
• of matches
Differences of 30\ or more
77 (66\)
11 of matches 65 (71\)
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
216 Valerie M. Hudson
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.
Scripting International Power Dramas 217
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.
Scripting International Power Dramas 219
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.
220 Valerie M. Hudson
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
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.
[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
224 Brian L. Job and Douglas Johnson
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
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."
third section the specifics of its application concerning the United States
and the Dominican Republic are detailed.
( CONTROL J
_,_..
_,----'-' --------\;_,--------------------
', .------'=:._ _ _ _ _--....
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.
Representation
Figure 9.2: Three Examples of the PAL Semantic Network Representation of Events
in the UNCLESAM Model
Balaguer
Event-228 -Threaten<
OR-Government
Escalate-Action-Suppress - - - -
---Opposftion
Balaguer threatens to have the government increase fts suppression of the opposftion.
Ideology
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.
NOTE:
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:
NOTE:
••• indicates response produced by UNCLESAM model
+++ indicates posture change, an mput to UNCLESAM
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
A Rule-Based Model of U.S. Foreign Policy Making 239
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
242 Brian L. Job and Douglas Johnson
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.
244 Brian L. Job and Douglas Johnson
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
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.
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:
[--- 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
Foreign Policy Decision Making as Knowledge-Based Reasoning 253
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.
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.
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
256 D. A. Sylvan, A. Goe/, and B. Chandrasekaran
Classification Classification
Module I Module 11
Threats to lnternationa
Japanese Problems
Energy Supply 1 Facing Japan
Security
Classification
Module m
Plan
Selection
Module
'
Energy-related Policies
which perform the generic task of classification, while the fourth performs
the generic task of plan selection.
FlowOueTo MajorChangeln
Change In ExportCapability
ExportPolicy
Immediate Cost
Major Cost
Foreign Policy Decision Making as Knowledge-Based Reasoning 259
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
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
260 D. A. Sylvan, A. Goe/, and B. Chandrasekaran
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.
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.
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.
Foreign Policy Decision Making as Knowledge-Based Reasoning 265
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.
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.
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:
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.
270 D. A. Sylvan, A. Goe/, and B. Chandrasekaran
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.
Foreign Policy Decision Making as Knowledge-Based Reasoning 271
[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.
Foreign Policy Decision Making as Knowledge-Based Reasoning 273
[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
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
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:
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-
280 Margee M. Ensign and Warren R. Phi/lips
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.
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
282 Margee M. Ensign and Warren R. Phi/lips
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.
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.
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.
A "Glass Box" Approach to Representation 287
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.
288 Margee M. Ensign and Warren R. Phi/lips
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
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.
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.
"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
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
The Expertise of the Senate Foreign Relations Committee 299
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.
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
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.
The Expertise of the Senate Foreign Relations Committee 301
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.
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
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.
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
The Expertise of the Senate Foreign Relations Committee 305
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.
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.
The Expertise of the Senate Foreign Relations Committee 307
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.
The Expertise of the Senate Foreign Relations Committee 309
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
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
Reproduction of Perception in the Early Cold War 313
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)
r---
build defend build world prudent
socialism socialism capitalism dominatio n restraint
- - +------- -------·- -----
------l------~
Round:
build gain. 1against
social ism support II
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.
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_~-----------
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
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
Reproduction of Perception in the Early Cold War 321
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
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.
Reproduction of Perception in the Early Cold War 323
/
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
Reproduction of Perception in the Early Cold War 325
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
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
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
336 D. J. Sylvan, S. J. Majeski, and J. L. Milliken
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.
Theoretical Categories and Data Construction 337
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.
338 D. J. Sylvan, S. J. Majeski, and J. L. Milliken
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.
Theoretical Categories and Data Construction 339
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.
340 D. J. Sylvan, S. J. Majeski, and J. L. Milliken
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,
McNamara, and Rusk, 3/6.
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,
McNamara, and Rusk, 3/6.
[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.
Theoretical Categories and Data Construction 341
***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.
342 D. J. Sylvan, S. J. Majeski, and J. L. Milliken
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.
Theoretical Categories and Data Construction 343
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
344 D. J. Sylvan, S. J. Majeski, and J. L. Milliken
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.
Theoretical Categories and Data Construction 345
(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.
346 D. J. Sylvan, S. J. Majeski, and J. L. Milliken
[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
347
348 John C. Mallery
Lexical-Interpretive Semantics
Semantic perception is the process of mapping from a syntactic rep-
resentation into a semantic representation (Mallery and Duffy, 1990). Tra-
Semantic Content Analysis 351
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
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
Semantic Content Analysis 353
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:
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.
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.
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)
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
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
Semantic Content Analysis 361
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
Semantic Content Analysis 363
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
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.
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.
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
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.
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:
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.
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.
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
Semantic Content Analysis 375
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
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
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
Semantic Content Analysis 379
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
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.
Semantic Content Analysis 385
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.
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
Temporal Representation
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
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 = = = = = =
c D
I I disjoin Is
after
overlaps
superiors
I
I
B X
I E
A
F
disjoints
before
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
415-817
~ ~316-942
42-382
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
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 ...,
- ---
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 .
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
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
Time Space: Representing Historical Time 403
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
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.
Time Space: Representing Historical Time 405
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.
407
INDEX
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