Political Science:
Artificial Intelligence Applications

Gavan Duffy
and
Seth A. Tucker

Syracuse University
Maxwell School of Citizenship and Public Affairs
Department of Political Science
Syracuse, New York 13244
(315)443--2416
gavan@mailbox.syr.edu
satucker@mailbox.syr.edu

reprinted from Social Science Computing Review
Spring, 1995
copyright, 1995 Duke University Press
All Rights Reserved



Q: What is the difference between an artificial intelligence program and a computer program?

A: About $25,000.




This joke perhaps overstates the case. Nevertheless, it contains a kernel of truth, which accounts for its wide circulation in artificial intelligence laboratories in the early 1980s. Especially in the popular and trade presses, enthusiasts wildly oversold artificial intelligence (henceforth AI) as a technical panacea, attracting speculative investment but simultaneously creating unreasonable expectations. AI programmers, upon whose shoulders the weight of these expectations ultimately rest, underscore with their little joke the irony of this folly. In the final analysis AI programs are just computer programs --- nothing more and nothing less.

Yet AI is a special kind of computer programming in that it endeavors to emulate aspects of human intelligence, e.g. reasoning, perception, planning, problem—solving, and language understanding. For Patrick Winston (1992: 5), this emulation distinguishes AI from computer science, while AI's concern with computation distinguishes it from psychology. Yet the former distinction is not all that clear. Even the computer pioneers of the early postwar period viewed their primitive devices as mental prostheses. And many of the computational techniques considered cutting—edge AI only 15 years ago today routinely appear in off—the—shelf programs widely disseminated to desktops throughout the world. When it succeeds, AI becomes computer science.

If the distinction between AI and not—AI is fuzzy, distinctions between different classes of AI programs are doubly so. In reviewing the main AI applications in political science, we confess our inability to categorize these efforts neatly. We can provide only a loose categorization, founded more on family resemblances than any formal set of necessary and sufficient conditions for category membership. Individual readers may wish to quibble with our category scheme, but we doubt their ability to construct one not similarly problematic.

We do not claim to present an exhaustive survey. Many worthy applications have surely escaped our net. We have tried, however, to recount what our sensors detect as the more important or interesting applications. Likewise, owing to the fuzziness of the distinction between AI and other forms of computer modeling, we have avoided any conservative demarcation. We prefer to err on the side of including applications some might not consider AI. We limit ourselves to applications directed toward empirical research in political science. Many systems have been developed that have practical applications in government, but these lie beyond our purview here.


Production Systems

Popular primarily for the comparative ease with which one can construct models and also for their ostensible ability to capture rational decision processes, production systems are the most prevalent AI form applied in political science. Production systems are collections of ``if.....then'' rules. When the conditions specified in the antecedent(s) of any rule obtain, its consequent(s) fire. Consequents typically posit new conditions that may in turn trigger other rules. In some production systems, these rules also include (subjective or objective) probabilities that govern whether consequents fire in the presence of the antecedents.

Rule—firing in production systems, or``forward—chaining,'' can be considered a species of deductive inference. In fact, a rule—firing is simply an instance of modus ponens, while chains of rule—firings exemplify the hypothetical syllogism. Production systems can also ``backward—chain,'' or traverse rule chains from a terminal consequent to the conditions that might have caused it. Backward—chaining is a species of abductive inference. In the context of deductive logic, this form of reasoning is considered fallacious (affirming the consequent). However, this fallible form of reasoning can be used to produce theory—based hypotheses (Peirce, 1955: 150-156; Charniak and McDermott, 1985: 453-456), where the body of rules constitute the theory. In production systems, backward—chaining over a trace of actual rule—firings also generates useful explanations of the system's output behavior.

In traditional computer programs, conditional expressions are ``hard—coded,'' or written into programs as procedures. Production systems, however, treat these conditionals as data. This practice makes production systems far more flexible than traditional computer programs. In ``expert systems'' or ``knowledge—based systems,'' a family of applications of production systems, persons trained in the use of a production system (``knowledge engineers'') assist persons with expertise in some substantive domain (e.g., medicine, law, mass spectography, automotive mechanics) as they input rules (heuristics) pertinent to that domain. Thus, a single program (or ``expert system shell'') can model expert decision—making in a wide variety of fields.

Production systems invariably focus on very restricted knowledge domains. In general, as the number of rules in a production system increase, the amount of time necessary to evaluate their various conditions grows disproportionately. The degree of disproportion is largely a function of the similarity of conditions across rules, the number of disjunctive conditions, and the technical skill with which the production system is programmed. Nevertheless, this combinatorial explosion limits the scope of substantive political questions that production systems can address. Political modelers must constrain their domains by imposing simplifying assumptions so stringent that their models no longer represent adequately the complex political processes they purport to describe (cf. Mallery, 1987).

This problem for political applications of production systems parallels the dilemma that faces those who apply microeconomic rational choice models to political problems. Some production system modelers, drawing upon Milton Friedman's (1953) defense of such models, profess unconcern for process validity so long as they attain outcome validity. That is, they argue that their models need only produce results that correspond to empirical political outcomes, and not necessarily also to processes that produce those outcomes. Critics would of course contend that unconcern for process validity frees modelers to construct whatever model they wish to produce the desired result. On their view, explanations based on such models might well be entirely vacuous. These problems aside, production systems offer researchers a tool with which to model more complex decision—making processes than the paper—and—pencil methods that formal theorists generally employ. Combinatorial difficulties arise much earlier in the latter than the former. Consequently, production systems afford modelers the opportunity to introduce greater numbers of actors, choice alternatives, nuance, and context. Since production systems can generate explanations of their behavior using backward—chaining, this added complexity in no way detracts from the modeler's ability to explain the model's performance.


The Cuban Missile Crisis

Thorson and Sylvan (1982) present a production system model that exemplifies this advance over more traditional formal modeling techniques. Their JFK/CUBA production system models American decision—making during the Cuban missile crisis. While earlier formal models characterized these decisions in terms of utility payoffs, Thorson and Sylvan explore more deeply. They use JFK/CUBA to examine the inferential processes leading to President Kennedy's decisions during the crisis. Thorson and Sylvan use production rules to represent the contents of Kennedy's beliefs regarding the Soviet Union, military policy, negotiations, etc. They also employ productions to incorporate contextual information regarding the state of the conflict, which Thorson and Sylvan term ``ambient information structures.''

While one may dispute its description of Kennedy's knowledge, the real value of the JFK/CUBA model lies in its treatment of context. Thorson's and Sylvan's analysis suggests that alternative understandings of Soviet intentions produce radically different decision outcomes. Perhaps more importantly, Thorson and Sylvan show how analysts can systematically manipulate these and other contextual factors to produce counterfactual scenarios. By systematically comparing the outcomes of these scenarios to those of the actual scenario, analysts can draw conclusions regarding the relative importance of particular contextual factors in producing observed outcomes.


Political and Economic Development

Phillips and Ensign (1982) model apply AI researcher Chuck Rieger's production system to analyze governmental decision—making regarding political and economic development. Their production rules simulate several archetypical decision—makers, each representing an alternative development strategy. By manipulating initial levels of input variables, such political instability and agricultural production, their model indicates that more traditional development strategies produced less instability than did strategies suggested by modernization theory. Ensign's (1985) companion study models the decison-making of international bankers in determining whether to grant credit to underdeveloped states.


Military Expenditures Decision—Making

Majeski (1989) offers MEDM, a rule—based model of military expenditure decision—making. While acknowledging the value of rational choice analyses, he believes that attention to process validity will ultimately produce more coherent, accurate, and useful understandings of decision—making processes. In particular, he argues that production systems can more adequately capture the boundedness of decision—makers' rationality.

MEDM includes of two sets of rules. The first are heuristics that summarize the worldviews for each of the groups involved in the process (the three services, the President, and the Congress). These rules categorize and otherwise interpret variables representing situational exigencies. These interpretations are then passed to a second set of rules representing bureaucratic maxims for success, and decisions are output.


Hudson's Role—Theoretic Model

Hudson (1991) describes a production system that embodies an account similar to Majeski's. She draws on sociological role theory to argue that the positions of individuals within the social structure strongly condition their behavioral expectations. Behavior that might seem to deviate rational courses of action may in fact prove rational once the individual's role is taken into account. Roles change across individuals, of course, and these changes may be responsible for any individual's non—conformity to the prescriptions of rational choice models that fail to account for their effects. Hudson's production model attempts to capture the influence of roles on decisions. The merit of both Majeski's and Hudson's models resides in their use of production rules to model heuristics and maxims decision—makers employ to reduce their information costs. A critic might wish to dispute the contents of these rules, but it would be incumbent upon her to show how different, and arguably more accurate, rule sets would produce simulated decisions that correspond better to actual decisions. By this logic, however, it might also prove instructive to compare systematically the decision outcomes that these models produce to those generated by rational choice models.


UNCLESAM and the Dominican Republic

With their UNCLESAM model, Job and Johnson (1991) offer a relatively straightforward application of production system technology to political analysis. UNCLESAM is designed to represent a decision-maker with a limited knowledge repertoire of facts and values. When presented with data describing conditions surrounding the Dominican Republic from 1959 to 1965, UNCLESAM invokes productions representing heuristics believed to capture the rules of U.S. interaction with the Dominican Republic during this period.

Job and Johnson contend that much political action can be explained as rule—following of this sort, and on this basis believe production systems an appropriate modeling vehicle. One could argue to the contrary, however, that actual policy—making is not nearly as monological as this and most other production system models portray it. Policies emerge from the multilogical, discursive interplay of competing, organized interests, and no single set of decision rules can capture that interaction.


Integrating Mathematical Models

Anderson and Thorson (1982) discuss a hybrid model that combines a production system model of government decision—making and a difference equation model of flows of commodities and populations. The simulated Saudi government, in reponse to user—declared input describing the current state of the international system, alters equation parameters to reach stated production goals.

This effort demonstrates an important potential application area for AI technology. The integration of AI decision—making models might dramatically improve the performance of quantitative models of processes in which humans intervene. AI techniques might incorporate the qualitative effects of human intervention in a much subtler and more nuanced fashion than can be achieved by standard methods, such as the use of dummy variables. Moreover, filtering quantitative process models with AI belief models would support interesting counterfactual analyses. It would enable investigations, for instance, of the likely effects upon economic performance of subtle changes in the beliefs of economic policymakers.


Japanese Energy Security

Sylvan, Goel, and Chandrasekaran (1990; 1991) take a useful step in this direction with JESSE, the Japanese Energy Supply Security Expert. Their model incorporates multiple decision—makers as well as more of the cognitive, interpretive processes that enter into their decisions. more obvious political actors involved in decision-making as well as the cognitive aspects of the process. Thus, they seek to achieve some measure of process validity as well as outcome validity.

JESSE queries users about the nature of the present energy situation vis--'a—vis Japan. From responses to these queries, JESSE constructs a knowledge base describing the Japanese energy security position. Consulting this knowledge as well as a set of production rules, JESSE proposes policies drawn from a stored set of plans. The JESSE model clearly outpaces other political applications of production systems in terms of technical sophistication. Likewise, its employment of multiple institutional actors in the policy recommendation process is commendable. Nevertheless, the inflexibility of its built—in categorization schemes for representing knowledge is open to criticism. The categories reflect the coder's worldview, which may not coincide with the worldviews of actual decision—makers.

One might also criticize JESSE's use of a pre—specified repertoire of plans. Some of the learning systems described below induce productions from datasets representing decisionmakers' experiences. A valuable contribution could be made by marrying these learning techniques to JESSE's substrate or to some other serious production system.


Logic Programming

Logic programming in AI generally employs the technique of resolution theorem proving. In this approach, axioms in first—order predicate logic are first restated in a normal form. Next, the conclusion is negated and similarly restated in normal form. Iteratively using the disjunctive syllogism, the theorem prover attempts to derive a contradiction. If the introduction of a negated conclusion among the axioms produces a contradiction, the conclusion itself must be true if the axioms are true. Winston (1992: 293-303) provides a clear summary of the resolution technique, which forms the basis of the PROLOG computer language.


Social Reproduction

Banerjee (1986) constructs a PROLOG model to demonstrate how, consonant with Immanuel Wallerstein's world systems theory, political organizations reproduce themselves within a larger society. Borrowing Jean Piaget's action schemata to represent the cognitive reinforcement of organizations by their ability to explain and interact with environmental phenomena, Banerjee simulates the reasoning of ideal—typical members of several societies. The model shows which organizations reproduce schemata and which do not.

Whether one accepts or rejects the accounts Banerjee deploys to reconstruct these social arrangements and his use of Piagetian theory, his creative use of a PROLOG model stands as useful counterpoint to the relatively simple, and sometimes simplistic, applications of production systems technology. By their nature, production systems can describe only a relatively small subset of logical relationships.


Conceptual Intelligibility

Conceiving politics as a set of processes that require the continual evaluation of essentially novel events, Bennett and Thorson (1991) ask how political actors render such novel events intelligible to themselves. They contend that this categorization, framing, or interpretation of events crucially affects actors' decisions. Thus, they criticize the assumption, plainly apparent in most production systems models, that decision—makers automatically render events intelligible by matching their features to a stored set of templates. They claim that this practice effectively denies the creative problem—formulating activities of human decision—makers.

Sylvan and Thorson (1992) draw upon a similar argument to criticize their own JFK/CUBA production system model. On the understanding of Kennedy's problem representation they had coded into JFK/CUBA, the model rejected as implausible Dean Rusk's view that the Soviets would likely retaliate with military action against West Berlin for any American airstrikes in Cuba. Because Rusk viewed the American dilemma regarding Cuba as essentially symmetrical to the Soviet quandary regarding West Berlin, he could reasonably infer Soviet retaliation on West Berlin. Since the productions used to represent Kennedy's problem representation included no such symmetry, it ruled out Rusk's scenario.

To exemplify the argument, Bennett and Thorson construct alternative models of the deterrence concept, while Sylvan and Thorson construct alternative models of the concept of offensiveness in US—Soviet interactions. In each paper, the authors discuss the action consequences of each alternative understanding. They present their models as code fragments written in Scheme, a dialect of the Lisp programming language used primarily for instructional purposes. As Lisp and its various dialects implement the lambda calculus, we class these models under the logic programming rubric. These models might usefully be recast in PROLOG, or in some other suitable language, to produce working models of event interpretation under alternative conceptual renderings of deterrence.

Bennett and Thorson illustrate a point that suggests a vital task for AI applications in political science. The methods political modelers generally use are extensional --- they presuppose that the meanings of concepts are non—problematical and operate over the similarly non—problematic instances of those concepts. The meanings of many political concepts like deterrence, however, vary across political actors and even political analysts. This variation diminishes the persuasive force of extensional analyses involving such concepts. Readers who cannot agree to the analyst's meaning or who believe that the analyst's meaning does not capture the actors' meanings will find the analysis too tenuously linked to political reality.

These papers illustrate the importance of developing dialogical, or even multilogical, models of foreign policy decisionmaking. In real life, different players contribute different --- sometimes widely different --- understandings of the decision situation. Models that represent only one understanding of the situation cannot hope to capture the essence of the decision. Whether cast as a production system or in game—theoretic terms, political decision models must somehow represent the discursive processes by which decisionmakers mutually attenuate their individual understandings of their dilemmas.


Belief Models

Although one can argue that production systems can model belief, they capture a relatively small part of belief --- the ``knowing how.'' They do not capture the ``knowing that.'' Belief models typically capture both. They typically include a semantic network for representing declarative knowledge plus a set of productions for representing procedural knowledge. Belief models are consequently more general and flexible than simple production systems. Moreover, as Thorson (1984) observed, production systems inadequately capture decisionmaking processes if they fail to incorporate the effects on decisions of the pre—existing beliefs of decisionmakers.

Semantic networks are complex data structures that describe relationships between and among general concepts and particular instances. A useful way to conceptualize such networks is to consider them collections of nodes. Some nodes represent the subjects of relations, some represent the objects of relations, and some represent the relations themselves. In well—designed semantic networks, the relations can themselves be the subjects or objects of other relations.

Most belief models link universal nodes, or concepts, to nodes that represent particular instances of those concepts. The concept nodes ``subsume'' those instance nodes, meaning that the latter inherit any relationships in which the former participate. For example, if a network represents a `support' relation, with the conceptual subject `Stalinist' and object `wars of national liberation' and if `Stalinist' also subsumes the particular `Brezhnev,' then the network automatically also represents the belief that Brezhnev supports wars of national liberation.


The Ideology Machine

Abelson and Carroll (1965) present the first political belief model, and perhaps the earliest application of AI in political inquiry. Their model represents the contents of a political ideology and responds to queries about foreign policy from the perspective of that ideology. References to this system occasionally refer to it as ``the Goldwater Machine'' because Abelson and Carroll early on used it to model the beliefs of Barry Goldwater. However, Abelson and Carroll present it as a more generic interpreter of ideological beliefs and refer to it as ``the Ideology Machine.''

In a later essay, Abelson (1973) uses the Ideology Machine to explore the human tendency to interpose simplified symbols between themselves and the ``external'' world The model consists of a vocabulary, conceptual categories, episodes, and, most importantly, ``master scripts'' representing ideological perspectives. Abelson suggests that these master scripts guide the processing of political information, influencing in large measure the responses of ideologues to political events.

>From the perspective of the succeeding generation of research in political psychology, this early effort today seems quaint and even naive. From the perspective of modern AI research, the search procedures of the Ideology Machine seem unnecessarily inefficient. As an importany contributor to both literatures, Abelson himself would probably agree. Nevertheless, as the first computer simulation of political belief, the Ideology Machine stands as an important achievement. It demonstrated that one can model the core beliefs of political actors computationally and use them to generate predicted responses to policy problems automatically.

Abelson soon began a long collaboration with AI researcher Roger Schank. Together, they published a work describing Conceptual Dependency (CD) as a computational theory of human cognition (Schank and Abelson, 1977). Abelson devotes a large proportion of his 1973 essay to recasting it in terms of that theory.

Scripts play an important role in CD, as does the notion of ``semantic primitives.'' Here, relationships that one would express in natural language as verbs are reduced to molecular structures of atomic primitives. For instance, if X were to give Z to Y, the CD approach would represent a ATRANS (transfer of possession) of Z from X to Y. If, however, X were to donate Z to Y. X and (possibly) Y would also experience increases in a mental state scale. Although the list of semantic primitives has expanded greatly since their introduction, the approach supports the automatic identification of meaning equivalences by providing a category scheme that expresses the very large number of expressible relationships to combinations of a relatively small set of primitives.


Politics

Carbonell's (1978) POLITICS can be viewed as an updated implementation of the Ideology Machine. It makes full use of the CD framework and addresses several of the research problems arising from the Ideology Machine research. Consulting relatively detailed descriptions of ideological beliefs, POLITICS generates a wide variety of goal—directed inferences from a political conflict event description. These include predictions of subsequent events and suggestions for American policy planning. Building upon the success of POLITICS, Carbonell (1981) extends the model to explore planning under conditions of uncertainty. Rarely, if ever, do political decisionmakers formulate action strategies without counterplanning against possible contingencies. They must, for example, consider the possibility that opponents might ally with others to enhance their response capacity. They must likewise also consider their own alliance opportunities and evaluate likely opportunities for strategic compromise.

In extending POLITICS to incorporate counterplanning, Carbonell essentially suggests a method for modeling strategic interaction. The formal, game—theoretic methods often used for this purpose necessarily describe strategic interaction very abstractly. By simulating planning and counterplanning in belief models, political analysts may be able to devise models of strategic interaction that capture the contents of strategy and not merely their form. Serious political analyses conducted within such an environment might well generate more content—sensitive and context—sensitive analyses of strategic situations without many of the simplifying but potentially validity—threatening assumptions of formal techniques.


The Policy Arguer

With his Policy Arguer, or POLI, Taber (1992) constructs a knowledge base that represents the ideological perspectives common to isolationists, militant anti—communists, and pragmatic anti—communists of the 1950s. Given an input event and the selection of an ideological perspective, POLI interprets the meaning of the event in accordance with the ideology. On that basis, it then forward—chains through a set of productions representing procedural beliefs shared across ideologies, producing a set of policy recommendations consonant with the ideological interpretation of the event.

Taber takes an important step toward providing a political application with a multilogical capability. Still, more could be done. Although it contains multiple perspectives, POLI produces arguments that depend only upon one of those perspectives. POLI might usefully be extended to allow those differing perspectives to debate one another over appropriate courses of action.

Taber and Timpone (1994) describe a more general version of the model in greater detail. Expressing their concern for process as well as outcome validity, they describe a model consistent with Anderson's (1983) ACT* framework. Anderson presents ACT* not merely as an AI system for modeling beliefs but also as a general theory of cognition. Any claim of process validity for POLI thus depends crucially upon Anderson's claim that belief models of the ACT* class adequately capture human cognition. This latter claim is open to dispute.

In particular, both ACT* and POLI strictly separate declarative knowledge (or the contents of semantic memory) from procedural knowledge (or the contents of production memory). The problems with this separation are unlikely to be noticed until one asks the system to learn concepts about its productions, since learning would occur in the declarative memory. This difficulty might not affect POLI as it currently stands, but it might well arise were POLI asked to construct taxonomies of productions or to invent productions on the fly for particular situations by analogizing from the productions that would fire in analogous situations. In any event, in the absence of a cognitive—theoretical motivation for the separation between declarative and production memory, Anderson's cognitive claims and Taber's and Timpone's process validity claims should be viewed with some skepticism.



Computational Hermeneutics

Researchers in computational hermeneutics (Mallery, Hurwitz and Duffy, 1987) seek to marry AI natural language understanding technology with hermeneutic text interpretation. They hope to provide techniques for simulating text interpretations. By altering the simulated ``preunderstandings'' upon which readers necessarily rely when interpreting any text, they hope to test hypotheses concerning the inferential effects of particular preunderstandings upon the readings of particular texts (Duffy, 1994). This approach bears an obvious affinity to the counterfactual analysis of Sylvan's and Thorson's JFK/CUBA and of Taber's POLI, both of which are discussed above.

The idea for computational hermeneutics arose from an interest in ``precedent logics'' regarding political action decisions. On this view, historically—conditioned decisions, like those in international politics, can be explained neither as myopic updated expectations nor as blind rule—following behavior. Researchers in precedent logics believe political actors to be endowed with far richer senses of history. From salient precedents, or relative exemplars drawn from their understandings of history, they construct both the rules they follow and the outcomes they expect. Precedent construction, on this view, becomes an integral part of the policy formulation process, and involves the hermeneutic construction of historically rich preunderstandings upon which to ground contemporaneous interpretations of political reality.

The process of constructing precedents involves transforming naked events into chains of antecedents and consequences. The factors comprising events, and the events themselves, are selected and linked together by causal and intentional connectives to yield a possible ordering or configuring such that each contributes to the possibility or necessity of the next. Plot—like in structure, these interpretations assign roles and rationales that, along with situational factors, are assembled into unfolding sequences of choice, accident, and error --- all conditioned by circumstances (Alker, Bennett, and Mefford, 1980: 194-195).

From the earliest work on precedent logics (Alker and Christensen, 1972; Alker and Greenberg, 1976), the affinity between this substantive idea and the emerging AI modeling techniques was plain. Alker (1975) first conceived the modeling part of the effort as ``formal hermeneutics.'' However, the effort being so much more directed toward empirical content than to abstract form, ``computational hermeneutics'' has since became the established term.

Research proceeeded along what Duffy (1994) calls the ``macrostructural path,'' in which substantive analyses were conducted using available technologies. Published computational analyses along this path include Tanaka's (1984) model of Chinese political decision—making and the Alker, Lehnert, and Schneider (1985) analysis of Toynbee's account of the Jesus myth. Non—computational analyses of narrative policy constructions continued along a parallel track. Two of the more noteworthy are Alker's (1988) analysis of Thucydides' Melian Dialogue using Nicholas Rescher's argumentation formalism and Mefford's (1992) account of the changing metaphors in Soviet policymakers' rhetoric through the period of perestroika and glastnost

. While some researchers pursued macrostructural research, Mallery and Duffy, following an important early statement by Winograd (1980), noted the affinity between precedent logics and the implementational research of Winston and Katz (Winston, 1982; Katz and Winston, 1982; Katz, 1988). Using simplified plot outlines of Shakespearean plays and descriptions of physical objects, the programs of Katz and Winston automatically discovered precedents from textual inputs. Mallery and Duffy developed a new system --- Relatus --- capable of performing similar tasks using much larger textual inputs. A lay description of Relatus appears in Alker, Duffy, Hurwitz, and Mallery (1991), while the technical details of various components appear elsewhere (Duffy and Mallery, 1984; Mallery, 1985; 1985a, Duffy, 1986; Duffy, 1991; Mallery, 1991).

Relatus parses English—language text and represents its contents in a referentially—integrated semantic network. All analyses of textual contents, including precedent detection and lexical categorization (Mallery, 1991), occur within this network environment. Most analyses rely upon the ``constraint—interpreting reference'' operations devised by Mallery (1985a). For instance, when mapping text contents from parsed syntactic representations, procedures walk the parse graph posting constraints on its nodes. These are then gathered to form a plan for mapping the sentence's contents gracefully (i.e., maintaining coreferences across sentences) into the contents already represented in the network (cf. Duffy and Mallery, 1984). Constraints are similarly interpreted when finding precedents. A procedure called ``semantic inversion'' converts semantic network structures into constraints capable of finding replicas and near—replicas of those contents in the same or in another semantic network. Constraint—interpreting procedures are designed to speed the search of existing contents by reasoning about the order in which constraints are interpreted (cf. Mallery, 1985a).

Mallery and Duffy conceive of Relatus as a substrate in which all varieties of computational hermeneutic analysis may be conducted. Several components, including a system for representing temporal knowledge, an activation system for automatically generating and applying rules, a lexicon system for acquiring and organizing multiple languages, and learning algorithms, exist in various stages of development but have not yet been integrated into the larger system. The system can spawn multiple ``belief systems,'' supporting simultaneous development of belief models for multiple actors. One research goal of the Relatus project would simulate policy debates between multiple actors.

In originally identifying ``formal hermeneutics'' as a prospective method of political analysis, Alker (1975) drew upon Schank's and Abelson's CD framework for illustrative purposes. The CD reduction of terms to ``semantic primitives'' presupposes a strict synonymy of terms and analyticity of sentences for which Quine (1953) provides a devastating refutation. For this reason, Mallery and Duffy adopt a ``lexicalist'' approach that models textual contents using the terms that appear in the text. They support the classification of lexical categories (Mallery, 1991) for the purposes of analyzing textual contents. However, they do not force a particular category scheme into the computer representation of textual contents, as do Schank and Abelson. The strength of the CD approach lies in its ability to detect meaning equivalences. But Quine's refutation of analyticity implies that the CD reduction to primitives can introduce a great deal of error by eliminating linguistic nuance. Nevertheless, the detection of meaning similarities is a prime task for any natural language understanding program, and some category scheme must be employed to support it. The lexicalist Relatus model finesses the difficulty by allowing users to employ any category scheme --- and even multiple category schemes --- consonant with their analytic interests.

Because the Relatus project resembles basic research in AI more than any other political science application, the research to date focuses more on developing the substrate and less on preparing application demonstrations prematurely. See Mallery (1991) for a list of the early Relatus applications. See Duffy (1994) for a statement on the role of Relatus for research in precedent logics and computational hermeneutics.


Machine Learning

Computer scientists have developed several methods for machine learning, many of which have received political application. Machine learning technology has advanced a great deal over the past generation, partly due to the declining costs of memory. Because particular instances must be scanned and rescanned a great many times in order to induce useful generalizations, machine learning is computationally intensive. As costs continue to decline, and as parallel hardware becomes more widely diffused, machine learning techniques are likely to appeal to political modelers and other applications programmers as a useful data—reduction technique. As political modelers gain experience with machine learning techniques, other applications will likely emerge.


Levenshtein Distance Metrics

Schrodt (1991) employs the notion of Levenshtein distance to predict whether an interstate event sequence will result in war. Levenshtein distance measures the conceptual distance between two sequences as the weighted sum of the operations --- insertion, deletion, and replacement --- needed to transform one sequence into the other.

Weights are generally assigned on the basis of the relative importance of the type of each element in the sequence. Schrodt derives the weights inductively. Taking a small training set from the Behavioral Correlates of War (BCOW) dataset, he first assigns small weights to clearly inconsequential element types and large weights to clearly consequential ones. He then applies a procedure that finds the weights for the other types that (a) maximize the distance between sequences that end in wars from those that do not and (b) minimize the distance between sequences within each of the two groups. Using these weights, he computes the Levenshtein distance of event sequences in a test set also drawn from the BCOW dataset. The procedure performs remarkably well, correctly predicting the outcomes (war or non—war) in each event sequence of the test set.


Genetic Algorithms

Genetic algorithms refer to systems that learn by emulating processes of natural selection: mutation, environmental fitness, etc. Holland (1975) classifiers, for example, create taxonomies by arbitrarily mating data elements, then checking environmental feedbacks to see whether this mutation can survives. The fittest mutations (those that encompass most observations) survive, while unfit mutations become extinct. The system continues this process iteratively until it achieves a prespecified level of stability. Although genetic algorithms might be misused as technical support for atheoretical fishing expeditions, they offer researchers the opportunity to discover regularities in their data that they may not have anticipated.

Schrodt (1989) applies Holland classifiers to the COPDAB (Conflict and Peace Databank) descriptions of American and European interactions from 1948 to 1978, inducing a set of production rules. He then applies these rules to 40--day event sequences to generate predicted behaviors for the next 20 days. Schrodt reports that the model predicts short—term behavior as accurately as statistical best estimator techniques.

Cohen (1992) uses Holland classifiers to test the specification of his maximum likelihood model survey respondents' candidate preferences. Using respondents' issue preferences, knowledgeability, and demographic features as explanatory variables, he finds that the classifiers and maximum likelihood model make substantially identical predictions. From this he can conclude that, using these data, no other model specification would provide a better account of respondents' candidate preferences.


Inductive Interaction Detection

Sherman, Mallery, Unseld, and Duffy (1992; Mallery and Sherman, 1993) apply Inductive Interaction Detection, or I2D, to induce a typology of event sequences from the SherFACS international conflict management dataset (Sherman, 1994). I2D, developed by Unseld and Mallery (1991), extends J. Ross Quillian's learning algorithm so that in can operate over non—rectangular datasets of phase—structured event sequences, such as SherFACS. Nodes in the I2D typologies describe condition—action rules and enumerate the cases that conform to the rule.

Sherman et al. represent the SherFACS data in Mallery's (1994) Feature Vector Editor, a data structure that combines dataset and codebook as a unit to facilitate data entry and manipulation. I2D operates on nodes in this representation. Feature Vector nodes representing cases and subsequences of cases are also integrated into Duffy's (1994) Timebase program. By supporting constrained rule induction within particular time intervals, Timebase allows Sherman et al. to compare rules induced within particular historical eras to test whether the rules had changed across those eras.

The I2D research and Schrodt's applications of Levenshtein distance and genetic algorithms support the program in precedent logics discussed above. Each induce generalizations from international events data to identify clusters of precedents upon which policymakers might draw for decision guidance. Schrodt begins the process of validating induced precedents by employing them in short—term predictions. Other methods of validation would be helpful, however. In particular, if induced precedents actually provide policy guidance, they should appear in policymakers' internal deliberations.


Interstate Conflict Simulations

In recent years, renewed interest has emerged in computer simulations of international political conflict. Although these efforts arguably do not qualify as AI, the fact that they simulate state decision—making qualifies them under our loose understanding of AI. While these simulations fall generally within the class of models that describe the social macrobehavior that emerges from the interactions of actors endowed with micromotives (Schelling, 1971; 1978: 137-166; Schrodt, 1981), they descend more directly from the Bremer and Mihalka (1977) simulations of interstate conflict. Bremer and Mihalka create a world of 98 contiguous hexagons, arranged in a pattern similar to chicken coop fencing. In this model, predatory states with varying degrees of inaccurate perception attack one another where they see some likely advantages. States forge offensive and defensive alliances, and, on the basis of war outcomes, divide territorial and power spoils. While Bremer and Mihalka do not conduct systematic tests, they report a tendency for the simulation to produce universal empires, in which one state eventually achieves global domination.


Exploring Realpolitick

Cusack and Stoll (1990) extend the Bremer—Mihalka model in several ways. Among them, they add several new state decision procedures to the ``primitive power—seekers'' of Bremer and Mihalka. These include rational actors, collective security seekers, and power balancers. By varying these state types in hundreds of simulation runs, they conduct systematic tests of the effects of these decision rules on the survival and endurance of states and state systems.

Cusack and Stoll also add a component that simulates empire maintenance and civil war. In simulation runs that include this component, states that pay insufficient attention to paying the costs of maintaining their empires risk civil war. If the rebels succeed, empires can decompose into several new states that then enter into interstate conflicts. Most importantly, however, Cusack and Stoll use their system to test systematically the competing theoretical propositions of several schools within the ``realist'' school of international politics scholarship.


Adding Concurrency

Duffy (1992) challenges the ``serial assumption'' he finds implicit in the Cusack and Stoll simulation models. Primarily because they implement their models on serial devices, Cusack and Stoll restrict their simulations such that only one war can occur at any one time and only one alliance can be formed at any one time. Duffy implements a similar model on a massively parallel Connection Machine, representing each state as a computing device that operates concurrently with other states. Multiple wars can occur and multiple alliances can form concurrently, further constraining the action possibilities of states.

Comparing his statistical analyses of state and system survival and endurance with those of Cusack and Stoll, Duffy finds that his parallel implementation produces many of the results evident in the serial model. An important difference regarding state decision—making emerged, however. Where Cusack and Stoll found that acting on utility—maximizing principles always tended to enchance the survival and endurance of states, Duffy found this to be the case only to the extent that other states in the environment also acted on the bases of those principles.


Conclusion

Political science and artificial intelligence are both ``sciences of the artificial,'' in the apt phrase of political scientist and AI pioneer Herbert Simon (1981). They both concern themselves with artifice, or intentional effects of human actors. Techniques appropriate for modeling natural processes may not always be fully appropriate for modeling processes involving beings quite capable of manipulating them intentionally. The effects of the various notions, traditions, and meanings of human actors have long served to buttress the skepticism of many political scholars to the model construction and analysis activities of colleagues who pattern their research activities more after the natural scientists than after the humanists.

Indeed, those who apply AI to political problems very much seek to expand the scope, acceptability, and relevance of modeling in political science by constructing models that incorporate the effects of those human notions, traditions, and meanings. They seek to humanize political models. But they freely admit that they have only just begun their enterprise. They recognize that they must enlarge the community of scholars engaged in the effort if it is to succeed. They do so by encouraging students and colleagues to familiarize themselves with AI techniques and to seek collaborative projects with their AI colleagues across campus.

We have sought in this essay, then, to provide something more than a simple literature review. We hope that we have imparted a sense of the variety and innovative spirit that characterizes the nascent tradition we describe. We have also indicated, where appropriate, where additional contributions might effect further advances. Finally, we hope that our accounts communicate our sense that AI technology offers the prospect of more compelling descriptive foundations for political analysis.


References


Abelson, Robert P. 1973. ``The Structure of Belief Systems.'' In Roger Schank and Kenneth M. Colby, eds., Computer Models of Thought and Language. San Francisco: W. H. Freeman, pp. 297-340.

Abelson, Robert P., and J. D. Carroll. 1965. ``Computer Simulation of Individual Belief Systems.'' American Behavioral Scientist. 8: 24-30.

Alker, Hayward R. Jr. 1975. ``Polimetrics: Its Descriptive Foundations.'' In Fred Greenstein and Nelson Polsby, eds., Handbook of Political Science. v. 7. Reading, MA: Addison—Wesley, pp. 140-210.

Alker, Hayward R. Jr. 1988. ``The Dialectical Logic of Thucydides' Melian Doalogue.'' American Political Science Review. 82: 805-820.

Alker, Hayward R. Jr., James P. Bennett, and Dwain Mefford. 1980. ``Generalized Precedent Logics for Resolving Insecurity Dilemmas.'' International Interactions. 7:165-206.

Alker, Hayward R. Jr., and Cheryl Christensen. 1972. ``From Causal Modelling to Artificial Intelligence: The Evolution of a UN Peace-Making Simulation.'' In J. LaPonce and P. Smoker, eds. Experimentation and Simulation in Political Science. Toronto: University of Toronto Press, pp. 177-223.

Alker, Hayward R. Jr., and William J. Greenberg. 1976. ``On Simulating Collective Security Regime Alternatives.'' In G. Matthew Bonham and Michael J. Shapiro, eds., Thought and Action in Foreign Policy. Basel: Birkhauser Verlag.

Alker, Hayward R. Jr., Gavan Duffy, Roger Hurwitz, and John C. Mallery. 1991. ``Text Modeling for International Politics: A Tourist's Guide to Relatus.'' In Valerie M. Hudson, ed., Artificial Intelligence and International Politics. Boulder: Westview, pp. 97-126.

Alker, Hayward R. Jr., Wendy C. Lehnert, and Daniel K. Schneider. 1985. ``Two Reinterpretations of Toynbee's Jesus: Explorations in Computational Hermeneutics.'' In Graziella Tonfoni, ed. Artificial Intelligence and Text Understanding. Special issue of Quaderni di Ricerca Linguistica. 6: 49-94.

Anderson, John R. 1983. The Architecture of Cognition. Cambridge: Harvard University Press.

Anderson, Paul A., and Stuart J. Thorson. 1982. ``Artificial Intelligence Based Simulations of Foreign Policy Decsion Making.'' Behavioral Science. 27: 176-193.

Banerjee, Sanjoy. 1986. ``Reproduction of Social Structures: An Artificial Intelligence Model.'' Journal of Conflict Resolution. 30: 221-252.

Bennett, James P. and Stuart J. Thorson. 1991. ``Reasoning and Intelligibility.'' In Valerie M. Hudson, ed., Artificial Intelligence and International Politics. Boulder: Westview, pp. 127-148.

Carbonell, Jaime G., Jr. 1978. ``POLITICS: Automated Ideological Reasoning.'' Cognitive Science. 2:27-51.

Carbonell, Jaime G., Jr. 1981. ``Counterplanning: A Strategy—Based Model of Adversary Planning in Real—World Situations.'' Artificial Intelligence. 16: 295-329. Charniak, Eugene, and Drew McDermott. 1985. Introduction to Artificial Intelligence. Reading: Addison—Wesley.

Duffy, Gavan. 1991. ``Time Space: Representing Historical Time for Efficient Event Retrieval.'' In Valerie M. Hudson, ed., Artificial Intelligence and International Politics. Boulder: Westview, pp. 386-405.

Duffy, Gavan. 1986. ``Categorial Disambiguation.'' In AAAI-86: Proceedings of the American Association for Artificial Intelligence. Los Altos, CA: William Kaufmann, Inc., pp. 1057-1061.

Duffy, Gavan. 1994. ``Events and Versions: Reconstructing Event Data Analysis.'' International Interactions. 20: 147-167.

Duffy, Gavan, and John C. Mallery. 1984. ``Referential Determinism and Computational Efficiency: Posting Constraints from Deep Structure.'' In AAAI-84: Proceedings of the American Association for Artificial Intelligence. Los Altos, CA: William Kaufmann, Inc., pp. 101-105.

Ensign, Margee M. 1985. An Anlysis of Private Bank Loans to Developing Countries: The Relationship Between Bankers' Images and Their Lending Policies. New York: Gordon and Breech.

Friedman, Milton. 1953. ``The Methodology of Positive Economics.'' in Essays on Positive Economics. Chicago: University of Chicago Press, pp. 3-43.

Holland, John H. 1975. Adaptation in Natural and Artificial Systems. Ann Arbor: University of Michigan Press.

Hudson, Valerie M. 1991. ``Scripting International Power Dramas: A Model of Situational Predisposition.'' In Valerie M. Hudson, ed., Artificial Intelligence and International Politics. Boulder: Westview Press, pp. 194-220.

Katz, Boris. 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, pp. 55-77.

Katz, Boris, and Patrick H. Winston. 1982. ``A Two—Way Natural Language Interface.'' In Integrated Interactive Computer Systems. In P. Degano and E. Sandawall, eds. Amsterdam: North-Holland.

Majeski, Stephen. 1989. ``A Rule Based Model of the United States Military Expenditure Decision—Making Process.'' International Interactions. 15: 129-154.

Mallery, John C. 1985. ``Universality and Individuality: The Interaction of Noun Phrase Determiners in Copular Clauses.'' In Proceedings of the 23rd Annual Meeting of the Association for Computational Linguistics. pp. 35-42.

Mallery, John C. 1985a. ``Constraint—Interpreting Reference,'' Cambridge: MIT AI Laboratory Memo 827.

Mallery, John C. 1987. Thinking about Foreign Policy: Finding an Appropriate Role for Artificially Intelligent Computers. SM Thesis. MIT Department of Political Science.

Mallery, John C. 1991. ``Semantic Content Analysis: A New Methodology for the Relatus Natural Language Environment.'' In Valerie M. Hudson, ed., Artificial Intelligence and International Politics. Boulder: Westview, pp. 347-385.

Mallery, John C. 1994. ``Beyond Correlation: Bringing Artificial Intelligence to Events Data.'' International Interactions. 20: 101-145.

Mallery, John C., and Frank L. Sherman. 1993. ``Learning Historical Rules of Major Power Intervention in the Post—War International System.'' Paper presented to the International Studies Association. Acapulco.

Mansbridge, Jane J. 1990. Beyond Self—Interest. Chicago: University of Chicago Press.

Mefford, Dwain. 1992. ``From Struggle to Dialogue: The Revolution in Soviet Political Language.'' Paper presented to the Annual Meeting of the International Studies Association, Atlanta.

Peirce, Charles S. 1955. ``Abduction and Induction.'' In Justus Buchler, ed. Philosophical Writings of Peirce. New York: Dover.

Phillips, Warren R., and Margee M. Ensign. 1982. ``Self—Reliance and Political Development: Simulations of Inner Environments.'' International Political Sciene Review. 3: 455-478.

Quine, Willard V. 0. 1953. ``Two Dogmas of Empiricism.'' In From a Logical Point of View. New York: Harper and Row, pp. 20-46.

Schank, Roger C., and Robert P. Abelson. 1977. Scripts, Plans, Goals and Understanding. Hillsdale, NJ: Lawrence Erlbaum Associates.

Schrodt, Philip A. 1989. ``Short—Term Prediction of International Events Using a Holland Classifier.'' Mathematical and Computer Modelling. 12: 589-600.

Schrodt, Philip A. 1991. ``Pattern Recognition of International Event Sequences: A Machine Learning Approach.'' In Valerie M. Hudson, ed., Artificial Intelligence and International Politics. Boulder: Westview Press, pp. 169-193.

Sherman, Frank L. 1994. ``SherFACS: A Cross—Paradigm, Hierarchical and Contextually Sensitive Conflict Data Set.'' International Interactions. 20: 79-100.

Sherman, Frank L., John C. Mallery, Sigrid C. Unseld, and Gavan Duffy. 1992. `` Grammars of Conflict and Cooperation in the Post—War International System.'' Paper presented to the International Studies Association. Atlanta.

Simon, Herbert A. 1981. The Sciences of the Artificial. 2d ed. Cambridge: MIT Press. Sylvan, Donald A., Ashok Goel, and B. Chandrasekaran. 1990. ``Analyzing Political Decision Making from an Information—Processing Perspective: JESSE.'' American Journal of Political Science. 34: 74-123.

Sylvan, Donald A., Ashok Goel, and B. Chandrasekaran. 1991. ``Modeling Foreign Policy Decision Making as Knowledge—Based Reasoning.'' In Valerie M. Hudson, ed., Artificial Intelligence and International Politics. Boulder: Westview Press, pp. 245--273.

Sylvan, Donald A., and Stuart J. Thorson. 1992. ``Ontologies, Problem Representation, and the Cuban Missile Crisis.'' Journal of Conflict Resolution. 36: 709-732.

Taber, Charles S. 1992. ``POLI: An Expert System Model of U.S. Foreign Policy Belief Systems.'' American Political Science Review. 86: 888-904.

Taber, Charles S., and Richard Timpone. 1994. ``The Policy Arguer: The Architecture of an Expert System.'' Social Science Computer Review. 12: 1-25.

Tanaka, Akihiko. 1984. ``China, China Watching, and CHINA_WATCHER.'' In Donald A. Sylvan and Steve Chan, eds., Foreign Policy Decision Making: Perception, Cognition, and Artificial Intelligence. New York: Praeger.

Thorson, Stuart J. 1984. ``Intentional Inferencing on Foreign Policy: An AI approach.'' In Donald A. Sylvan and Steve Chan, eds., Foreign Policy Decision Making: Perception, Cognition, and Artificial Intelligence. New York: Praeger, pp. 280-309.

Unseld, Sigrid C., and John C. Mallery. 1991. ``Inductive Interaction Detection for Complex Datamodels.'' MIT Artificial Intelligence Laboratory Memo 1298.

Winston, Patrick H. 1982. ``Learning New Principles from Precedents and Exercises.'' Artificial Intelligence. 19: 321-350.

Winston, Patrick H. 1992. Artificial Intelligence. 3d ed. Reading, MA: Addison—Wesley. Winograd, Terry. 1980. ``What Does It Mean to Understand Natural Language.'' Cognitive Science. 4: 209-241.


Gavan Duffy is Associate Professor of Political Science in the Maxwell School of Citizenship and Public Affairs of Syracuse University. He is the author of several AI programs in the areas of text understanding, lexical representation, and temporal knowledge representation.

Seth A. Tucker is a doctoral candidate in Political Science in the Maxwell School of Citizenship and Public Affairs of Syracuse University. His forthcoming dissertation will describe a computational model of strategic negotiations.

Click Here to Return to Top Page