Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.
6 Meso-Level Formal Models I n this chapter we describe and discuss formal models of human Âbehavior at a level of aggregation and detail between the micro and macro levels. Such models are often referred to as meso-level models. Typically the models represent interactions and influences among individuals in groups and cover both individual and group phenomena and their interactions. These models include several voting and social decision models, social net- work models, link analysis, and agent-based modeling (ABM). The models have been developed in varied disciplines, including social psychology, sociology, anthropology, economics, and computer and communications sciences. Voting and Social Decision Models Understanding and predicting social phenomena requires good models of individuals and groups. The behavior of a group can differ from that of the individuals that comprise it. A science of aggregation is needed to model the behavior and actions of collections of people. There is a need to know how individual beliefs, goals, and skills combine on various tasks, such as problem solving and decision making. This section covers voting models that assume people reveal their true preferences, game theory models that assume people vote strategically, and social psychological models that con- sider how individual preferences might change in a group setting. 215
216 BEHAVIORAL MODELING AND SIMULATION What Are Voting Models? The research and models from voting theory provide a natural place to begin an investigation into aggregation for both pragmatic and conceptual reasons. Governments, terrorist groups, and alliances all make decisions by âvoting.â Some follow formal voting rules and procedures and others informally aggregate competing desires. Thus, our use of the term âvot- ingâ goes well beyond formal (e.g., electoral) registering of a preference to much less formal situations in which a preference is exercised or a decision is made with input from multiple individuals. Conceptually, voting models are valuable for three reasons: (1) a sub- stantial body of theory exists, (2) that theory shows no shortage of counter- intuitive results, thus highlighting the challenges of aggregation, and (3)Â the theory highlights a key point: to model groups well, one must be able to model individuals and the interactions between them. State of the Art in Social Decision Modeling We first describe the basics of preference theory. We then discuss results from social choice theory that reveal the problems created by aggregation as well as briefly comment on game theoretic models of strategic voting. The distinction between social choice theory and game theory hinges on behavioral assumptions. Social choice theory assumes that people truthfully reveal their preferences. Game theory does not. It assumes that people act strategically, which may or may not lead them to reveal their true prefer- ences. The game theory models also enable one to understand how and why various institutional rules matter. We also discuss research from psychology that addresses how choices are made in a group context. Preference Theory Preferences capture how much people value or desire things. They dif- fer from choices, which are what people select. Modelers define preferences over a set of alternatives. These alternatives can be outcomes, or they can be policies that produce outcomes (Page, 2007). Preferences impose an ordering over the alternatives. It is customary to write the preferences of someone who prefers apples (A) to bananas (B) as follows: A > B. Most modelers make two assumptions about individual preferences: that a person can compare any two alternatives (completeness) and that a person does not exhibit any preference cycles or internal contradictions (transitivity). â e W might have alternatively considered models of riots or collective ecosystem mainte- nance, but the related literature is not as deep or well thought out.
MESO-LEVEL FORMAL MODELS 217 If a person claimed to prefer apples (A) to bananas (B), and bananas to coconuts (C), and then claimed to prefer coconuts to apples, one might think that person was irrational. Formally, it would be said that the person exhibits a preference cycle in which A > B > C, but C > A. When individual preferences satisfy both completeness and transitivity (i.e., A > B > C and A > C), then they are called rational. If a person has rational preferences and if the modeler rules out indif- ference, then that personâs preferences can be written as an ordered list from the most to the least preferred alternative. Given a set of five alternatives, A, B, C, D, and E, one personâs preferences might be written A > B > C > D > E, and another personâs might be written E > D > C > B > A. This construction does not represent strengths of preferences. One per- son might strongly prefer A to B and strongly prefer B to C. Another person might have the same preference ordering but strongly prefer A to B and only weakly prefer B to C. To capture these relative strengths, one can assign pay- offs or utilities to each alternative. Payoffs are not considered here because comparing these utilities across people is considered a dubious practice. Social Choice Theory If the members of a group have identical preferences, then aggregating those preferences is straightforward. One can think of the group as one big individualâand for some groups that may not be a bad assumption. The aggregation of preferences becomes problematic when the group membersâ preferences are diverse. Preference diversity can be fundamental (people want different outcomes) or instrumental (people want the same outcomes but differ over the means to achieve them). In what follows, that distinction is ignored, but it becomes important when thinking about linking models. If voting models are to be linked with cognitive models, then the source of preference diversity is important to define because information can reduce instrumental preference diversity but has little effect on fundamental prefer- ence diversity. A collection of individuals with rational preferences may fail to have rational preferences as a group. We give an example and then state a gen- eral theorem. In this example, three military leaders have preferences over which city to use as a base of operations. The three candidate cities are Paris, London, and Berlin. The leaders are denoted L1, L2, and L3. Their preferences are as follows: Leader L1: Paris > London > Berlin Leader L2: London > Berlin > Paris Leader L3: Berlin > Paris > London
218 BEHAVIORAL MODELING AND SIMULATION Were these three leaders to vote on their choice between each pair of cities, London defeats Berlin two votes to one, Berlin defeats Paris two votes to one, and Paris defeats London two votes to one. Thus, the collec- tive preferences exhibit a cycle. Although the collective consists of rational individuals, the collective is not rational. In theoretical terms, the property of rationality does not aggregate. The possibility of a cycle is not an artifact of majority rule voting. Kenneth Arrow proved that any rule for aggregating preference orderings that is not a dictator produces cycles (Arrow, 1951). It requires only that preferences are diverse, rankings between two alternatives do not depend on a third irrelevant alternative, and rankings reflect unanimityâif every- one prefers A to B, then so does the collective. Arrowâs theorem does not imply that cycles are unavoidable, only that if one wants to avoid cycles, one has to sacrifice one of the other conditions of his claimâappoint a dictator, sacrifice unanimity, or violate indepen- dence of irrelevant alternatives. In general, as argued by Donald Saari, pref- erence cycles are more a function of the voting system than the voter. He suggests that voting paradoxes arise when the voting system fails to respect the natural cancellations of votes and so generates preference cycles (Saari, 2001). For example, one such voting system or scoring rule, the Borda rule (Marchant, 2000), does not create cycles. Under the Borda rule with three alternatives, a personâs top choice gets three points, her second gets two points, and her third gets only one point. Each alternative gets a score, making cycles impossible. Borda rule can, however, result in a tie, which is what would occur in the example of voting over cities. A tie isnât neces- sarily a bad thing. It reflects equal support for each alternative. Borda rule may thus seem to be better than majority rule, but we must keep Arrowâs theorem in mind. Borda rule must violate one of his conditions, and, in fact, Borda does not satisfy independence of irrelevant alternatives. In the example above, a fourth, irrelevant city could be introduced and change the outcome under Borda rule. The fact that by introducing irrelevant alterna- tives someone could change the outcome argues against using Borda rule. The debate thus moves from a discussion of the voter to a discussion of the scoring rules (Saari, 2006). Given that regardless of the voting rule individual agents may fail to reach a stable aggregation, organizational and institutional structures take on great importance. The rules for how a group makes decisions can have large effects on outcomes. For example, if someone has the power to set the agenda, then that person may have substantial power. Thus, even if an organization is democratic in principle, it may not be democratic in prac- tice, especially if one person controls the agenda.
MESO-LEVEL FORMAL MODELS 219 Strategic Voting In aggregating preferences, it can be assumed either that people vote sincerely or that they vote strategically. Strategic voting occurs even in large groups in mature democracies; people vote for candidates who they think can win rather than the candidates whom they most prefer. Alan Gibbard and Mark Satterthwaite have shown this incentive to misrepresent to be universal (e.g., Satterthwaite, 1975). Why does strategic voting further complicate matters? We have shown that rational individual preferences need not aggregate into a rational col- lective preference. Thus, it may not be possible to discern what a group would decide even if one knew the preferences of every member of that group. Given that people have incentives to misrepresent their preferences, they wouldnât reveal their true preferences anyway. Thus, one must discern how peopleâs true preferences get mapped into their actionsâin this case, their votes. And that requires a model of individual behavior in groups. The possibility of coalitions further complicates the analysis of vot- ing models. Subgroups may have an incentive to form a coalition to steer outcomes toward desired ends. This is seen in parliamentary systems, with Israel as an example. It may not be possible to predict which coalitions will form: politics makes strange bedfellows, and predicting those bedfellows can be difficult. In a group setting, social influence dynamics can muddy the picture even further, as people may change their preferences to align with the real or the inferred preferences of others. Concern for the preferences of others and for oneâs own standing in a group creates more indeterminacy in collective decisions. A striking example of such social influence effects is provided by the Abilene paradox, in which each person privately prefers X but believes that others prefer Z. If all group members revealed their true preferences, the group would clearly choose X. However, the desire to conform to what is (incorrectly) perceived to be the normative opinion can lead a member to suggest Z and others to agree (Harvey, 1974). While this counterintuitive outcome is probably rare in practice, it highlights the importance of realizing that social influence is not sim- ply a matter of one person seeking to change the preference of another. People also actively seek to align their preferences with important others. Both computational models and empirical studies have demonstrated that the impact of others on individual preferences tends to create uniformity of preferences among people who are closely connected. Dynamic social impact theory (LatanÃ©, 1996) predicts that people will change their prefer- ences to match those of others, with the impact based on both the strength (status) and the immediacy (social closeness) of others. The result is emerg- ing pockets of uniform attitudes based on social network clusters. Research
220 BEHAVIORAL MODELING AND SIMULATION on minority influence (e.g., Nemeth, 1986) also shows that the views of a cohesive minority, when clearly and consistently stated, can also change the opinions of majority members. Hence, knowing what the majority of individuals prefer at time 1 may not allow one to predict confidently what a group will choose at time 2. Relevance, Limitations, and Future Directions for Social Decision Models While voting models per se, especially those that compare specific vot- ing rules like Borda and majority rule, may seem more relevant to political science than to the situations that concern us here, the insights that can be drawn from these models are of critical importance. Two of the main rec- ommendations of this report are that modelers should recognize diversity of background, activity, and preferences and that they should embrace uncer- tainty. Nowhere does that advice ring more clearly and loudly than in under- standing the link from individual incentives to group behavior. Moreover, one can link diversity of background, activity, and preferences and uncer- tainty about all three into a general insight: the more diverse the members of a group in their general makeup (their background), their preferences, and their actions, the more uncertain one should be about their collective decisions and actions. For example, models that attempt to make predictions about the attitudes and behaviors of a group of nonÂcombatant civilians must consider the diversity of that group in terms of the sociocultural- e Â thnographic-economic background, preferences, and available actions. The more diverse the group on any of these three dimensions, the less certain the predictions. At a very practical level, the implication of recognizing diversity is to make the models more complex. Another practical implication is that model results should often be characterized in terms of how the diversity of the population being modeled impacts the results (e.g., show entropy or diversity indices to characterize the initial population and show how out- comes change as the initial population varies on this metric). Even if group models cannot be expected to make point predictions, they can provide a way to predict sets of possible outcomes. If one has even crude approximations of preferences, possible coalitions, and a set of possible voting rules, he can write game theoretic or agent-based models, and those models can provide some guidance for what might happen and, equally important, what probably will not happen. For a recent survey of these methods, see Kollman and Page (2006). The ability to apply voting theory depends on data, knowledge, and theory. For many of the problems relevant to this study, one would not have information about individual-level preferences. And, even if one did have access, the theory tells us that it is not possible to predict outcomes with certainty from that data. Equally important, one may not have knowledge
MESO-LEVEL FORMAL MODELS 221 of the voting rules. And, as discussed above, the voting rule has a substantial influence on the outcome. Thus, even with information about preferences, one would also need to know something about the process of preference aggregation in the group of interest. Finally, to apply these models, one needs models of how people behave in groups. Are the group members strategic? Do coalitions form, and who belongs to those coalitions? Imagine a model that includes the actions of a terrorist organization or of a nascent nation-state. One could make a black box assumption about how that organization or government acts. In other words, one could treat the group as an individual, presumably an individual who is the average of the group members. The voting models reveal the problems with that approach. Groups do not make choices as though they were a single individual. The natural way to improve the model would be to make the black box transparent and to allow for multiple characterizations of the collective decision-making pro- cesses that produce those outcomes. This will require data, knowledge, and models of the group of interest, but the potential payoff is large, as it will provide a more accurate assessment of the likely distribution of behaviors over the set of possible actions. Finally, empirical voting studies demonstrate that humans do not act in a strictly rational or strategic manner, hence calling into question the formal mathematical ârationalâ and âstrategicâ voting models. Summaries of the empirical literature point to the social rather than rational nature of vot- ing behavior; for example, people vote primarily along ethnocultural lines rather than according to their economic interests and display widespread voter ignorance (Friedman, 2005). As another example, research on the âvoter participation paradoxââin which it is asked why people vote at all, as each individual has virtually zero probability of affecting the out- come (Converse, 1964; Green and Shapiro, 1994)â both demonstrate this lack of rationality and suggests that there are huge variations in individual behavior. Turnout depends on a number of social factors, including the size of the electorate (as the size of the electorate grows, fewer voters turn out), the closeness of the competition (the closer it is, the higher the turnout), and the presence of an underdog (more turnout). This empirical work sug- gests both that the formal models and simplistic game theoretic models are inadequate and that the more detailed and nuanced behaviors possible in agent-based models (ABMs) are better at capturing the complexities of voting behavior. Social Network Models Networks are ubiquitous, and many techniques have been developed for analyzing, predicting, and understanding the world in terms of the set
222 BEHAVIORAL MODELING AND SIMULATION of connections among entitiesâa network. As the focus here is on social and behavioral modeling, we limit our discussion of network modeling techniques to those that have been and are being used to address individual, social, organizational, political, or cultural issues, rather than, say, gene interaction networks or computer networks. For a review of the field of net- work analysis, see Freeman (2004), and for the methodology, see ÂFreeman, White, and Romney (1991) and Wasserman and Faust (1994). What Are Social Network Models? Social network models view groups as consisting of a set of nodes (the members of the group) and a set of ties that connect them, which link together to form a network. The ties are often seen as pipes or roads along which various kinds of traffic flow, such as informational and material resources, as well as influences and coordination. Thus, a key aspect of net- work modeling is concerned with predicting (and controlling) what flows to whom at what time. Ties are also seen as providing a kind of underlying structure or topology that has effects on the performance of the group or individuals. A fundamental proposition of social network models is that a nodeâs position in the network (in conjunction with its attributes) deter- mines the opportunities for and constraints on action that it will encounter. A group-level corollary of this proposition is that the network structure of a group (together with other attributes of the group), determines the perfor- mance or outcomes of the group. Thus, network models differ from other models in placing less emphasis on characteristics of the nodes and more emphasis on the structure of connections between the nodes. Social network analysis (SNA) has received a great deal of attention since the terrorist attacks of September 11, 2001 (Borgatti and Foster, 2003). Phrases for fighting terrorism, such as âdisconnect the dotsâ and âit takes a network to fight a network,â and for doing business, such as âitâs not who you know but who or what who you know knowsâ and âare you networking?â have appealed to the imagination and raised awareness of this area. In addition, there have been successful applications of this approach. For example, social network information was used to locate Saddam Hussein, and several SNA tools have been used in various criminal investigations. Social network information is used in popular social net- working web services, like Friendster, to help students vet their dates. Traditionally, most SNA has focused on the analysis of relatively simple datasets involving a small number of social relations (often of just one kind) connecting a set of persons in some kind of group at a single point in time. Analysts in this area use computational techniques primarily to Âstatistically analyze these networks. This area has a long tradition, predating World War II. It emerged from the social sciences, particularly from social Âpsychology,
MESO-LEVEL FORMAL MODELS 223 anthropology, and sociology, and has now spread to organization science, economics, physics, and computer science. More recent work has focused on more complex networks involving large numbers of nodes of differing types (see section on Multimode Net- works below). For example, Carley (2003) has developed social network metrics that take into account not only relations among individuals, but also relations among tasks, relations among items of knowledge, assign- ments of tasks to individuals, relations of knowledge to individuals and other relationships. In addition, a key research interest today is in understanding network dynamics, both in the sense of how networks change over time (especially in response to attacks) and in the sense of how things flow over the network links. Carley (2003) has used multiagent models in a network context to predict and reason about change in social and other networks. State of the Art in Social Network Models In this section, we lay out the key concepts of SNA, starting with a discussion of the nature of the data and followed by an outline of the key analytical constructs, namely cohesion, centrality, equivalence, and cluster- ing. The section ends with a discussion of network evolution. Nodes and Ties The set of actors or agents that form the nodes of a network can consist of either individuals or collectives, such as organizations, cities, or countries. Nodes are assumed to possess characteristics that define their goals and affect their ability to achieve and exploit their network positions. These characteristics are modeled as a set of categorical and/or continuous attributes. In general, relations among nodes are modeled as dyadic 2-tuples (called ties, links, or edges) that bind exactly two nodes to each other. Therefore, a conversation among three people A, B, and C is typically modeled as three separate dyadic interactions consisting of A with B, B with C, and A with C. For the most part, the ties modeled among nodes typically belong to a general class known as social relations. These include such things as acquaintance (e.g., knows), kinship (e.g., brother of, father of), other social roles (e.g., friend of, teacher of), and affective relations (e.g., likes, dislikes). Each type of tie can be further characterized by relevant characteristics or attributes. For example, a friendship tie can be characterized in terms of intensity, closeness, and duration. In addition, network modelers often represent interactions over timeâ such as in-person meetings, communication, or fightingâas ties. Hence a tie
224 BEHAVIORAL MODELING AND SIMULATION is considered to exist between two nodes if at least one interaction between them is observed during a given period. The actual number of interactions may be recorded as an attribute of this tie. Interactions are inherently transitory and evanescent but are often seen as revealing the presence of underlying social relations. Interactions, such as conversations, provide the mechanism by which things flow through social relations, as when an actor transmits informa- tion to a friend through communication or when a person infects another with a disease via personal contact. Thus flows represent a third category of tie that a network modeler can choose to model. Typical flows of interest have been information, ideas, infections, material goods (such as guns and money), and such intangibles as energy and motivation. These are often referred to by network analysts as âtokens.â Multimode Networks When a categorical variable exists that distinguishes between different types of nodes, and, in addition, ties exist only between nodes of different types (and not within types), the resulting networks are referred as k-node networks or, in graph theory, as k-partite graphs, where k refers to the number of distinct types of nodes. These kinds of data typically arise in the context of recording affiliations between individuals and groups or events. For example, Davis, Gardner, and Gardner (1941) recorded which women attended which social events in a given season. Ties exist between women and events, but not among women and not among events. Similarly, it is common to record for each person in a group the organizations to which they belong(ed). And in organizational analysis, one can collect the number of hours that each person worked on various tasks or projects. Multinode networks can be analyzed directly or converted into simple 1-node networks by deriving co-occurrence indices. For example, a 2-node women-by-events network can be converted into a 1-node women-by- women network in which a tie between each pair of women is characterized by the number of events they attended in common. With multiple nodes, it is possible to represent the system as a whole as a meta-matrix (Carley, 2003). The meta-matrix is a conceptual device for identifying the set of networks within and among nodes of multiple classes. For example, given the three classes of nodesâpeople, knowledge, and activitiesâthe set of subnetworks possible is shown in Table 6-1. The second key concept is the entity ontologyâfor network analysis, this is the set of categories that defines the node classes and the link classes among the nodes used in a particular study. The table illustrates a particular ontology; other ontologies are needed for other applications.
MESO-LEVEL FORMAL MODELS 225 TABLE 6-1â Illustrative Meta-Matrix People Knowledge Activities People Social network Knowledge network Activity network Knowledge Information network Needs network Activities Precedence network Cohesion Models A fundamental concept in network modeling is cohesion. Cohesion refers to the connectedness or structural integrity of a network, and it is often interpreted in terms of the networkâs potential for coordinating among its members or exploiting knowledge that is distributed across the network. One aspect of network cohesion is density, which refers to the propor- tion of pairs of nodes that have a direct tie (i.e., are not dependent on an intermediary). A high density implies that, on average, each node is directly connected with many others. If the ties represent something like trust rela- tions, this indicates a group in which information can flow quite freely. Another aspect of cohesion is the average path distance, also known as characteristic path length. Path distance refers to the number of links in the shortest path between two nodes. A network with low average distance is one in which the lengths of the shortest paths between pairs of nodes are quite small, so that things flowing through the network can reach any or all nodes comparatively quickly. In the case of viruses or other infections, this is a measure of the vulnerability of the network to disease. In the case of the spread of best practices, it can be seen as a determinant of the potential performance of a continuously adapting system. Centrality Models A frequent analytical strategy in network modeling has been the iden- tification of key players who are disproportionately important due to their structural position in the network (Borgatti and Everett, 2006). The struc- tural importance of a node in a network is conceptualized as its centrality. One way to think about centrality is in terms of a nodeâs direct or indirect contribution to the cohesion or structural integrity of the network. For example, degree centrality is defined as the number of ties that a node has. If the total number of ties in the network is a measure of the cohesion of the network, then clearly degree centrality can be seen as each nodeâs âshareâ of the total cohesion. In this sense, the centrality measure implies a model of the sources of cohesion.
226 BEHAVIORAL MODELING AND SIMULATION Other well-known aspects of centrality include closeness centrality, betweenness centrality, and eigenvector centrality. If the graph theoretic dis- tance between two nodes in a network is defined as the length of the short- est path from one to the other, then closeness centrality is defined as the sum of distances from a node to all other nodes. To the extent that social ties among network members constitute pipes that transfer such Âtraffic as information or influence, closeness centrality gives the average time until the arrival of something flowing along the shortest paths. Betweenness cen- trality is the share of all shortest paths in the network that pass through a given node. High betweenness nodes are a kind of glue holding the network together; deleting nodes with high betweenness from a network tends to disconnect the network or make all paths much longer. Eigenvector cen- trality can be described, in simplified terms, as the extent to which a node is connected to many nodes that are themselves well-connectedâa kind of turbocharged version of degree centrality. Another way to think about centrality is in terms of the exploitability of a node position. This is the perspective taken by social capital theorists, who see a nodeâs position in the network as a kind of capital that the node can exploit for personal advancement or achievement. For example, a node with excellent closeness centrality is a short distance from other nodes in the network and is therefore well-positioned to hear information flowing through the network early, when it still confers a competitive advantage. A node with high betweenness centrality is in a position to make demands on others because these others need the central node in order to connect with others in an efficient manner. Finally, centrality can also be thought of as providing expected values for certain node outcomes in a particular flow process. For example, the formula that defines betweenness centrality gives exact estimates of the expected number of times that something flows over a node in a process in which tokens travel exclusively along shortest paths. Similarly, closeness centrality gives the expected values of the time to first arrival of a token flowing through a network, again using exclusively shortest paths. Degree centrality gives the frequency of arrival of a token in a process in which tokens travel along unrestricted random walks through the network. Thus, definitions of centrality carry with them a model of how things flow in a network. Equivalence Models Equivalence modeling refers to the branch of network modeling con- cerned with detecting nodes that play similar structural roles in the net- work (Borgatti and Everett, 1992). The simplest equivalence model is that of structural equivalence. A pair of nodes is structurally equivalent to the
MESO-LEVEL FORMAL MODELS 227 extent that they are connected (and not connected) to precisely the same third parties, regardless of whether they are tied to each other. Structurally equivalent nodes are structurally indistinguishable and substitutable. A fundamental claim in this kind of modeling is that, by virtue of being struc- turally isomorphic, structurally equivalent nodes will tend to have similar outcomes. Structural equivalence can also be seen as providing a formal definition for concepts of node environment and niche. Another equivalence model is called regular equivalence. This is a recursive model in which two nodes are regularly equivalent to the extent that they are connected to regularly equivalent third parties (but not neces- sarily the same third parties). Thus, two nodes do not have to have any contacts in common to be seen as equivalent, and indeed they can belong to entirely separate groups. As a result, the model can detect that both leaders of wholly unrelated organizations are playing the same role vis-Ã -vis their respective groups. Thus, it is a better model for the concept of social role than is structural equivalence. For example, given a network defined as the set of observed relationships among all people working in a hospital, regular equivalence can detect that two doctors of different patients are both playing the same role (i.e., doctor), whereas structural equivalence can detect only that two doctors of the same patients are playing the same role. The importance of regular equivalence is that it can discover latent or emergent social roles that have not been named and that the members of the network are themselves unaware of. However, the recursiveness of the definition, in which one needs to know the extent of regular equivalence between all other pairs of nodes in the network in order to calculate the regular equivalence of a given pair, makes this model computationally much more difficult than structural equivalence. Cohesive Subgroup Models An active area in network modeling is the identification of cohesive subsetsâdense regions of a network that have more ties within than to the rest of the networkâthat operate as units. The fundamental assumption in this work is that members of a cohesive subset will have more in common with each other than with nodes outside the subset (Borgatti, Everett, and Shirey, 1990). This occurs both because nodes with common attributes will tend to seek each other out, forming the cohesive subsets in the first place, and because members of cohesive subsets have disproportionate influence on each other, creating homogeneity within the group. Thus the h Â omogeneity of cohesive subsets results both from selection processes (simi- lar nodes joining together) and from influence processes (interacting nodes becoming similar to each other).
228 BEHAVIORAL MODELING AND SIMULATION Network Evolution Because social network modeling is a relatively young field that until recently had to fight for legitimacy, it is natural that it has concentrated on the impact of network variables on non-network or âtraditionalâ outcome variables, such as career success or team performance. However, as interest in networks has increased, so has research on the antecedents of network structureâin short, network evolutionâand the antecedents of a nodeâs position within network structures (e.g., why some nodes become more central than others) or of the emergent structural properties of the whole network (e.g., why some network structures are more robust than others). Empirical research has demonstrated several key factors that deter- mine who has ties with whom in a variety of networks. People who are physically near each other tend to communicate more, even in an age of asynchronous electronic communication. This effect of proximity (some- times called propinquity) is a special case of a more general principle known as Âhomophilyâthe tendency for individuals to have ties of various kinds with people who are like them on socially, culturally, or politically significant variables, such as geographic location, race, gender, age, social class, religion, culture, language, organizational affiliation, centrality, etc. Thus, these variables form the grist of most simulation models of net- work change. However, it should be noted that heterophilous mechanisms (in which opposites attract) also exist. Sexual relations, for example, are overwhelmingly heterophilous with respect to gender. In addition, most nonreciprocal relations such as âseeks advice fromâ or âgives orders toâ are Âheterophilous, so that less knowledgeable people seek advice from more knowledgeable people rather than from those equally knowledgeable. An important factor with elements of both homophily and heterophily is the activity focus. Common activities bring together people with similar interests, such as a bowling league or a political action group, creating homophilous linkages. However, as Alexis de Tocqueville noted as far back as 1835 (de Toqueville, 1835), these foci also tend to bring together people from different walks of life, creating heterophilous linkages across social boundaries. Another important factorânot unrelated to homophilyâis the tran- sitivity induced by such mechanisms as cognitive dissonance (Festinger, 1957) or balance (Heider, 1988). For example, if node A likes node B, and node B likes node C, then in many circumstances node A experiences some pressure to at least not dislike node C, thus increasing the probability of a tie forming between A and C. Finally, there are status-based mechanisms that are neither Âhomophilous nor heterophilous in which nodes are sorted by status, and all nodes pre- fer to interact with high-status nodes. In such cases, the high-status nodes
MESO-LEVEL FORMAL MODELS 229 exhibit homophily, because they prefer each other, while the low-status nodes exhibit heterophily, because they prefer high-status nodes. The model of preferential attachment developed to explain the pattern of which web- sites link to which other websites is a kind of status model. In preferential attachment, new websites link to existing websites with a probability pro- portional to the number of links the existing website already has, creating a situation in which, in terms of incoming ties, the rich get richer. In recent years, a number of researchers have modeled network change using simulation methods, typically variations of ABM (Zeggelink, 1994; Snijders, 2001; Carley, 2003). For example, in a stream of research she refers to as âdynamic network analysis,â Carley (2003) defines a subclass of ABMs that have multiple agents who dynamically form a network that evolves as the agents themselves learn and adapt. Agents take action on the basis of what they know, whom they know, and their own internal cogni- tive architecture, possible actions, and other factors. These models have been used to explore information diffusion, the impact of new technology, the evolution of networks, and the impact of interventions (e.g., to analyze the relative impact of different courses of action on terrorist groups). Key features are that there can easily be thousands of actors, with the exact number of actors limited only by storage space on the computer. The cognitive/Âcommunicative complexity of the model is limited by available computational capacity and processing time. Relevance, Limitations, and Future Directions Social network models and dynamic network analysis models can be used to identify key actors or groups. They are useful in understanding t Â errorist networks and in analyzing the criticality of nodes in those net- works. They are also useful in locating individuals: As mentioned, Saddam Hussein was located through an SNA of his contacts. Dynamic network analysis models can also be used to illustrate how the isolation of particular actors or groups will disrupt the flow of information or goods and services in both the short and long term dynamically. They can be used to show how groups or networks are likely to evolve under different conditions, techno- logical environments, etc. For example, the Construct model (a combined dynamic network and ABM) was used to contrast the effect of removing the top leader of al-Qaeda (bin Laden) and of Hamas (then Yassin) (Carley, 2004) and suggested that, for Hamas, performance would improve tem- porarily and the next leader would be Rantissi; in contrast, for al-Qaeda, performance would decrease and the next leader was indeterminate. In addition, they can be used to examine the impact of changes in recruit- ment on organizational performance, the effects of policing policies on civil unrest, the effect of technology and information sharing on organizational
230 BEHAVIORAL MODELING AND SIMULATION performance, the effect of detection technologies on information flow, etc. Dynamic network analysis models can also be used to create dynamic war- gaming scenarios by predicting the effects of courses of action on enemy and noncombatant behavior. Early work in social network modeling was mostly based on a branch of discrete mathematics known as graph theory. As a result, the models were fundamentally deterministic in character. These deterministic Âmodels do not lend themselves to prediction of populations (in which there tends to be a probabilistic distribution of behaviors and outcomes) nor of complex systems (see System Dynamics, Chapter 4). In these models, probabilistic thinking came into play only in relating deterministic variables to each other statistically. More recently, however, the field has begun to incor- porate stochastic thinking at a more fundamental level. For example, the exponential random graph models (also known as P* models) seek to model networks in terms of their latent tendencies to form micro structures, such as transitive triples or starlike subgraphs (called âmotifsâ in the physics literature) (Milo et al., 2002). By estimating a parameter for each kind of micro subgraph, the models can achieve a parsimonious description of the network in terms of a string of estimated parameter values, together with standard errors. This begins to make it possible to compare networks s Â tatistically with each other or with theory. Stochastic models also facilitate comparison of networks over time and indeed enable the estimation of rates of change in model parameters. In the long term, this line of work promises to yield continuous time models of network evolution, as opposed to current approaches to longitudinal analy- sis, which are limited to comparing snapshots of the network at discrete intervals in time. Similarly, most network models that are based on graph theory (and most are) are designed for binary data (i.e., a tie exists or it doesnât). Vali- dation of and extension of the metrics for nonbinary data is an ongoing research area that will eventually enable the capture of a wider range of social phenomena. Finally, most standard social network tools available on the web, in practice, are limited in their ability to handle more than 100,000 nodes with the exception of ORA (Carley et al., 2007). Visualization routines in gen- eral tend to be underdeveloped and work best with small datasets; however, for most military users, the goal is not to be able to visualize millions of nodes, but to have good preprocessing systems that subselect just the small portion of the network to view. From a military standpoint, many of the existing tools, because they are oriented around metrics, are too complex for the average soldier to use and contain little guidance on when to use which metric. Finally, much military data are multimode and multilink, and as a result they are cumbersome to process with most social network tools,
MESO-LEVEL FORMAL MODELS 231 with the exception of ORA. Rarely in military applications is it the case that the social network exists separate from, and needs to be assessed separately from, other types of networks, such as the activity network. All this being said, of the modeling tools described in this chapter, network models have had, to date, the biggest impact on military decision making. In particular, dynamic network tools that take into account the meta-matrix or that link to ABMs have been used to identify vulnerabili- ties in insurgent and terror networks, characterize political elites and track changes, identify local opinion leaders, and assess changes in beliefs and social influence. The most promising future directions involve linking these network approaches to other approaches, such as strategic reasoning Ã la game theory, or forecasting via ABMs, or geospatial identification by com- bining networks and map-based techniques. For example, placing network analysis in decision contexts enables reasoning about organizational change (Butts and Carley, 2006), while combining networks with spatial reasoning is facilitating analysis of the movement of terror groups to new locations of activity (Moon and Carley, 2007). Link Analysis Link analysis or link mining is related to SNA but has emerged as a distinct field centered on discovering patterns by looking at the relations among entities (see Getoor and Diehl, 2005, for a survey). Much of the work focuses on anomaly detection and link identification. What Is Link Analysis? Link analysis has emerged largely from computer science and forensics, with particular attention to work in machine learning. Historically, the term âlink analysisâ was used, particularly in the law enforcement area, to refer to approaches that let the analyst display and reason about the links between multiple types of nodes. Modern link analysis is a new subfield largely centered in computer sci- ence and statistics. Researchers and analysts in this area use computational techniques to locate patterns and subgroups based on a given set of infor- mation about paths, in which a path consists of a series of links that may connect nodes of different types, such as Joe + hamburgers + McDonaldâs. Extraction of links often requires massive data preprocessing or restruc- turing of databases (Goldberg and Wong, 1998). Given a set of paths, advanced data-processing techniques are combined with machine learning to enable rapid database transformation and pattern extraction. Key ques- tions often addressed are what paths are anomalies and what patterns can be inferred. Thus, much of the work in this area has focused on the iden-
232 BEHAVIORAL MODELING AND SIMULATION tification and recognition of patterns, data mining, and node identification and deidentification. Inferred patterns are then used to infer the âcauseâ of the pattern or to make predictions about future links. The main feature that distinguishes âsocial networksâ from âlink a Â nalysisâ in general and from âlink predictionâ in particular is the richness of the phenomena that are being explored and modeled. In social networks, the focus is often on producing qualitative and quantitative assessment about various items, such as leadership or influence or performance, test- ing, and estimation, whereas in link analysis the focus is on predicting quantities, such as the number of nurses who work at the hospital and give blood. Therefore, in social networks, the goal is to produce realistic models based on believed and theoretically grounded assumptions, a good descriptive model is looked for, and the researcher worries about how to do inferences. Parameter values encode semantics of interest in a specific application, and the research asks what the estimated values are and how much can one believe such estimates. It is only at the last stage that the social network Âtheorist worries about predictions. Good description is thought to yield good prediction. In contrast, in link analysis, the richness of a model is often sacrificed to statistical or computational efficiency. The prediction task, not necessarily the link prediction task, is the key focus. Accurate predictions and a quick black box that produces them are often viewed favorably in the literature. The analysis and comparison of various link analysis approaches are typi- cally weak; that is, little is done to compare the methods other than to com- pare speed. Furthermore, it is hard to make a case about why one should believe the guesses they produce other than on statistical grounds. In a sense there is no science that backs such predictions up, no theory for why these anomalies exist. Breiman (2001) discusses the differences between statistical and data mining approaches to analysis. These same differences apply to networks science (statistical) and link analysis (data mining). State of the Art Link analysis tools result in a mathematical representation of the rela- tion of different entities to each other vis-Ã -vis some problem. This math- ematical representation or âmodelâ of the underlying social behavior is discovered from the data and can then be utilized in other types of models, such as multiagent systems, to characterize a type of behavior. In modern link analysis, there are three fundamental concepts and two more general related concepts. First, there is the notion of similarity or distance among nodes. This distance is typically used to infer connectivity under the assumption that nodes that are similar or close will connect with other nodes in a similar fashion. Such a notion can derive from a formal
MESO-LEVEL FORMAL MODELS 233 (probabilistic) model or from other theoretical concerns and so is more deterministic. Link analysis algorithms can generally be categorized by their approach to similarity, or distance, and can be further divided by whether that distance is model based or algorithm based, and whether that distance is explicit or implicit. The second core concept is that of groups or clusters of elements of a network (typically of nodes) and the way both single elements and groups of them interact with one another. The idea of groups is central to most other link analysis papers. The third key concept is the link function, which translates similarity into the presence, absence, or weight of a link. Key differences in link analysis algorithms are often expressed in terms of dif- ferences in the link function. Less central ideas include types of nodes and links and, of course, time. In modern link analysis the analysis is done on the data itself, rather than on the network that has been inferred from the data. This avoids errors from the inference itself and from the relationship model that is being fitted. In addition, by assuming conditional independence of links, link analysts can leverage general statistical machinery. This provides an elegant way to deal with missing dataâany data one has are just a tiny snapshot of a rich distribution. Link analysis also deals easily with ârich links,â like multiparty links or multiple links between the same entities. In contrast, many of the social network tools have been developed with the âone dyad, one linkâ approach. In link analyses, the paths typically include nodes of multiple types, such as people and events and resources. In contrast, in a typical social net- work, the nodes are generally all of the same type or at most of two types. Each of the paths in link analysis is a single observation, hence temporal information on when a path occurred is available. In link analysis, no effort is made to take the paths and form the implicit networks. No assumptions are made about the completeness of the underlying network. In contrast, the social network modeler starts with a network and typically does not preserve path information, in the sense of information about observed tem- poral trajectories. Furthermore, social networks assume that the links are not independent events, whereas much of the work in link analysis assumes that each instance of a link is an independent event but subject to parameters that can be deduced. Finally, link analysis as a theory of anomaly detection is agnostic about the types of links and nodes that form the paths, whereas social network modeling has historically focused on networks in which the nodes are information-processing entities, such as people, organizations, or groups, and the links are the various factors by which they are connected, such as friendship, mentoring, financial transactions, or marriage. There are a growing number of link analysis tools, many of which are available on the web. Illustrative tools include GDA (Kubica et al., 2002;
234 BEHAVIORAL MODELING AND SIMULATION Kubica, Moore, and Schneider, 2003), PROXIMITY (Jensen and Neville, 2002), and PRMs (an extension of Bayesian Nets) (Getoor, Friedman, ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ Koller, and Taskar, 2001, 2002; Taskar, Abbeel, and ÂKoller, 2002; Taskar, Wong, and Koller, 2003)ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ . Common tools exist for doing a variety of tasks, including extracting of links from databases Â(Goldberg and Senator, 1998) and texts (Lee, 1998) and analysis of the extracted links (Chen and Lynch, 1992; Hauck, Atabakhsh, Ongvasith, Gupta, and Chen, 2002). Relevance, Limitations, and Future Directions Link analysis has widespread military applications in the creation of actionable intelligence from large diverse data sources and in the develop- ment of network modelsâsuch as terrorist network modelsâfrom partial and incomplete transaction data. Modern link analysis has focused on anomaly detection in large datasets. For many of the techniques that rely on machine learning, a key issue is having sufficient data with appropriate distributions so that the model can be âtrained.â In general, link analysis tools require a large quantity of labeled data that have been preextracted. A second key limitation of this approach is that many of the tools assume that the data exist in a file or database and cannot handle streaming data as they arrive or data that are âout of orderâ in terms of the entity classes. Third, many of the current nonproprietary methods do not scale well, greatly reducing the size of the datasets that can be handled. A fourth related limitation is that the models that are discovered when using a link analytic approach inherently assume that âtomorrow is like today.â Hence using these models to predict future behavior may be limiting. Fifth, to be useful, these models need to be expanded to handle streaming data. This is work in progress and Bayesian updating rules are being developed for an increasing number of models. A final limitation of link analysis is that the models that result, although fitting strong mathematically based theory, may not be reasonable from a social or behavioral perspective. For example, knowledge discovery rou- tines for finding groups will find groups that meet some predefined statisti- cal requirement; however, these groups may not match the definition of a group in empirically grounded social theory or even the everyday sense of what constitutes a group. While link analysis is generally useful for locating patterns and discern- ing structure, it is very limited in its ability to analyze downstream effects. For example, using link analysis to answer such questions as who has the most connections (degree centrality) can be done in only the most rudi- mentary way when links are viewed as independent. Consider the question, âIs Mustafa important because of the number of communications that go through him, or because his communications are the only ones connect-
MESO-LEVEL FORMAL MODELS 235 ing different sectors of a terrorist group?â Using SNA, this question can be addressed directly and easily with existing and well-Âvalidated metrics. When a link analysis approach is taken, in which links are viewed one at a time and treated as independent, a special-purpose and extremely complex model would need to be constructed. In principle, link analytic tools can be used to locate and construct the networks, and then social network or dynamic network metrics can be applied for predictive purposes. This is a promising direction for creating actionable intelligence. However, current link analytic models always use customized representations of the underlying network, making it difficult to transfer their results to other tools designed for prediction. A general- ized standard for representing the underlying network is needed. Advances in this direction include graphml and dynetml (Tsvetovat, Reminga, and Carley, 2004), which are XML languages for dealing with network data; however, both of these are insufficient to meet the needs for which link analysis results are used. An important issue for network modeling is the robustness of the models in the face of errors in the data. This is particularly an issue for hidden or stigmatized populations (e.g., criminals, terrorists) and for illicit or private relations (covert operations, political influence, etc.). To date, very little work has been done to assess the robustness of different network models in the face of different kinds of errors in the data, such as missing ties, missing nodes, gratuitous ties, and gratuitous nodes (as when a person who uses two different names is mistakenly entered as two different nodes). Similarly, few network models provide standard errors or confidence inter- vals for their outputs. Thus, it is simply unknown how much error in the data can be tolerated or whether a network model of flawed data does more harm than good. Another increasingly important issue for network modeling is the bounding of empirical networksâthat is, determining which nodes to include and which ones to exclude. Part of the conceptual base of network modeling is the interdependence of nodes. This creates a problem for artifi- cially bounding the networks that one wishes to model. One can arbitrarily choose to model the members of an organization or the residents of a vil- lage, but this does not stop the nodes from having ties with people outside the sample frame. To the extent that these unobserved ties affect what happens to the nodes, the models will fail to predict outcomes of interest. This problem cannot be eliminated, but it can be ameliorated by including larger chunks of the human network in the analysis, particularly chunks that correspond to natural boundaries. For example, if the computational and data collection issues can be overcome, modeling an entire village or other geopolitical unit is clearly preferable to arbitrarily modeling half of the village because of practical limitations. What is needed is investigation
236 BEHAVIORAL MODELING AND SIMULATION into the consequences of the different ways of bounding networks and into alternative ways of framing research issues to get around the boundary specification problem. A major area for future research is the study of models and algorithms to recover and/or discover link connectivity patterns, rather than node con- nectivity patterns (in the sense of Milo et al., 2002). This has potential for application, for example, to network privacy (reidentification, deidentifica- tion), to subgraph matching, and to motif discovery. Of primary impor- tance for real-world applications is the development of fast approximation algorithms that replicate the solution of successful algorithms for solving various problems, thus addressing the scalability issue that typically bur- dens algorithms that involve counting links in various ways. Another area that requires investigation is how to connect models of static and dynamic networks to observations and measurements. This is a general issue for all modeling techniques, and its implication for the broader impact of research is far-reaching and would include as a subtopic the integration of informa- tion from multiple sources, Ã la metamatrix, to support the discovery of interesting patterns. In addition, any simultaneous advances in automated data collection and computational algorithms for very large networks would significantly improve the usefulness of link analysis for the problems at hand. It is already possible to construct communication networks based on telephone logs, e-mails, etc. However, the degree to which one can infer different kinds of social relationsâsuch as trust, kinship, aid, conflict, etc.âfrom these data is still unknown, nor are alternative data currently available. Much of social network research has been based on survey research methodology, which is not applicable in the case of unwilling actors, such as enemies. Agent-Based Modeling of Social Systems The social and organizational sciences seek to understand not only how individuals behave but also how interactions among individuals generate macro-level outcomes. Understanding a social system requires more than understanding the individuals in it. It also requires understanding how the individuals interact with each other and how the results can be more than the sum of the parts. Agent-based modeling is well suited for this objective. â everal S research communities are currently exploring methodological approaches closely related to agent-based modeling under a variety of other names. Examples include Âmultiagent- based systems, agent-based computational economics, agent-based social simulation, multiÂ agent systems, and individual-based modeling. A sample of introductory readings from these various research communities can be accessed online at http://www.econ.iastate.edu/tesfatsi/ aintro.htm.
MESO-LEVEL FORMAL MODELS 237 What Is Agent-Based Modeling? Agent-based modeling is the computational study of systems that are complex in the following sense: (1) the systems are composed of multiple interacting entities and (2) the systems exhibit emergent propertiesâthat is, properties arising from entity interactions that cannot be deduced simply by averaging or summing the properties of the entities themselves. What distinguishes agent-based modeling from general complex sys- tems modeling, however, is the form of the entities that make up the system. A system can be complex even if its constituent entities are homogeneous units, such as CO2 molecules. In contrast, the constituent entities of an ABM are heterogeneous âagentsâ with internal states that can vary over time in response to internal deliberations as well as external forces, thus admitting the exploration of systems of heterogeneous agents with a range of social and learning capabilities. More precisely, the agents in an ABM can represent people (e.g., con- sumers, sellers, voters). They can also represent social groupings (e.g., f Â amilies, firms, communities, government agencies, nations), biological enti- ties (e.g., livestock, crops, forests), and even physical systems (e.g., weather, geography, transmission grids). When the interaction network formed by agents is contingent on past experience, and especially when the behaviors of agents in this interaction network continually adapt to past experiences, standard mathematical and statistical tools typically have only limited abil- ity to derive the dynamic consequences. In this case, agent-based modeling might be the only practical method of analysis. Agent-based modeling is a general-purpose technology. On one hand, the only constraints are the modelerâs purpose, imagination, and ability to encode. A modeler is free to make assumptions believed to be most relevant and realistic for an issue of interest. On the other hand, the realism of the resulting model will depend strongly on the extent to which the modelerâs assumptions are driven by data. In general, the more tightly a model has been constrained by real-world data, the smaller the space of possible outcomes. As detailed by Axelrod (1997, pp. 206â221), simulation in general, and agent-based modeling in particular, is a third way of doing science in addition to deduction and induction. Scientists use deduction to derive theorems from assumptions and induction to find patterns in empirical data. Simulation, like deduction, starts with a rigorously specified set of assumptions regarding an actual or proposed system of interest, but, unlike deduction, simulation does not prove theorems with generality. Instead, simulation generates data suitable for analysis by induction. In contrast to typical induction, however, the simulated data come from controlled experiments rather than from direct measurements of the real world. Con-
238 BEHAVIORAL MODELING AND SIMULATION sequently, simulation differs from standard deduction and induction in both its implementation and its goals. Simulation permits increased understand- ing of systems through controlled computational experiments. In particular, agent-based modeling can be used to investigate how macro-level effects and social behaviors arise from the micro processes of interactions among many agents. A general phenomenon exhibited by ABMs is large events. The logic of the central limit theorem states that the sum of a collection of random events produces a bell curve. In such cases, deviations from the mean, large or small, are rare. In ABMs, random effects can accumulate. These accumu- lations can be more than additive or even multiplicative. The result can be huge cascades: forest fires, riots, stock market crashes, epidemics, and even the collapse of governments. Moreover, ABMs can be used to estimate the probability of such extreme events (Gladwell, 2000). In summary, agent-based modeling applied to social, cultural, and orga- nizational processes uses concepts and tools from social science and com- puter science. It represents a methodological approach that could ultimately permit three important developments: (1) the rigorous testing, refinement, and extension of existing theories that have proved to be difficult to for- mulate and evaluate using standard mathematical and statistical tools; (2)Â a deeper, more integrated understanding of fundamental causal mechanisms in multiagent systems, whose study is currently hampered by artificial dis- ciplinary boundaries; and (3) a tool for exploration and evaluation of the potential impact of course of action and policy alternatives. State of the Art The goals pursued by ABM researchers take six general forms: empiri- cal description, empirical prediction, normative analysis, behavioral under- standing, heuristic understanding, and methodological advancement. Researchers pursuing empirical description ask: Why have particular macro-level structures and social behaviors evolved and persisted, even when there is little top-down control? Examples include trade networks, socially accepted monies, mutual cooperation based on reciprocity, and social norms. Agent-based modelers seek causal explanations grounded in the repeated interactions of agents operating in specified environments. In particular, they ask whether particular types of observed macro-level regu- larities can be reliably generated from particular types of ABMs. ABM researchers interested in empirical prediction ask: If this history of events were to take place, what would be the likely future consequences? These types of questions can be pursued in the context of ABM frameworks in which the modeler builds in scenarios of interest, introduces agents with
MESO-LEVEL FORMAL MODELS 239 realistic degrees of adaptability, and then tests to see how the agents react over time as the scenarios unfold. A third goal is normative analysis: How can ABMs be used as laborato- ries for the discovery of good rules of operation? ABM researchers pursuing this objective are interested in evaluating whether policies and institutional arrangements proposed for various types of social systems result in desir- able system performance over time. Examples include the design of auction systems, voting rules, and law enforcement practices. A fourth goal is the understanding of diverse behaviors. The perfor- mance of markets, democracies, and even traffic laws varies around the globe. We chalk up these differences to cultural or behavioral differences, but we lack a calculus of culture. We do not know whether or not slight variations in behavioral rules will result over time in widely divergent out- comes. ABM can help to illuminate the accumulation of effects from diverse behavioral rules and the extent to which slight variations in behavioral rules have substantial effects. This goal overlaps with both the empirical goal of prediction and the normative goal of good operational design, yet it is distinct from each. It necessitates a fundamental shift in how one looks at social systems. Standard models typically focus on the means of variablesâthe average expected outcomes. Yet often in agent-based modeling, the tail of the dis- tribution wags the dog, so to speak. For example, to predict the likelihood of a riot, what matters most is not the average level of civil unrest among a population but the percentage of people enraged enough to trigger a riot through disruptive behavior that others will then mimic. A fifth goal is heuristic understanding: How can greater insight be attained about the fundamental causal mechanisms in social systems? Even if the assumptions used to model a social system are simple, the con- sequences can be far from obvious if the system is composed of many interacting agents. The macro-level effects of interacting agents are often surprising because it can be hard to anticipate the full consequences of even simple forms of interaction. For example, one of the earliest and most elegant ABMsâthe city segregation (or âtippingâ) model developed by Nobel laureate Thomas Schelling (1978, pp. 147â155)âdemonstrates how residential segregation can emerge from individual choices even when everyone is fairly tolerant. A sixth goal is methodological advancement: How best can ABM researchers be provided with the methods and tools they need to under- take the rigorous study of social systems through controlled computational experiments? How best can they examine the compatibility of experi- mentally generated theories with real-world data? ABM researchers are exploring a variety of ways to address these issues, ranging from careful
240 BEHAVIORAL MODELING AND SIMULATION consideration of methodological principles to the practical development of programming, visualization, and validation tools. Perhaps the most provocative consequence of these methodological advancements is in the area of nonequilibrium science. Much of existing social science research, particularly research relating to organizations and institutions, is predicated on an assumption that systems are in equilib- rium. This allows one to compare modeled systems by the equilibria they implement. In contrast, the real world routinely exhibits a wide variety of nonequilibrium phenomena, such as abrupt transitions, crashes, and path dependencies. Agent-based modeling permits researchers to study out-of- equilibrium behaviors, hence it should ultimately help them to understand, evaluate, and characterize these phenomena (Arthur, 2006; Page, 2008). ABM Structural Properties ABMs can be structurally specified in widely diverse ways. Five dis- tinguishing structural properties of particular interest are as follows: the number of agents, the basic manner in which agents are represented, the cognitive sophistication of the agents, the social sophistication of the agents, and whether or not the agents are situated in a relational or spatial grid. Table 6-2 illustrates how these five structural properties differ across four classes of models currently used by ABM researchers: cognitive ABMs; dynamic network ABMs; cellular automaton ABMs; and rule-based ABMs. The table must be interpreted with some care. One caveat is that, in each model class, the actual level of realism depends on the degree to which agent attributes are based on actual data and the degree to which agent behavioral rules faithfully represent real-world processes. Another caveat is that, in principle, ABMs are ubiquitously applicable to problems that involve two or more agents whose behavior depends, at least in part, TABLE 6-2â Structural Differences Commonly Exhibited by Agent-Based Models Number of Agent Cognitive Social Grid Model Agents Representation Sophistication Sophistication Based Cognitive Few Rules High Low No Dynamic- Many Equations + Moderate High No network rules Cellular Few to many Equations or Low Low Yes automata rules Rule-based Few to many Rules Low Low Often
MESO-LEVEL FORMAL MODELS 241 on each other. Thus, differences commonly exhibited in current use do not necessarily reflect fundamental differences in capabilities. For example, the fact that cognitive ABMs currently tend to comprise relatively few highly sophisticated cognitive agents is due to processing power limitations and not to modeling or coding limitations per se. We now consider the ABM structural properties in greater depth. Number of Agents and Cognitive Sophistication As a general rule, the cognitive sophistication of the agents in an ABM is inversely proportional to the number of agents. On one hand, a model could comprise from 2 to 10 very cognitively sophisticated agents doing very in-depth knowledge-intensive tasks. In such a model, interactions among agents would typically be prescribed by protocols for interaction and by hierarchical precedents regarding who does what. Such models are more common in computer science and engineering; illustrative models are those involving BRAHMS, Soar, ACT-R, or Neural Networks (see Chapter 5). Models of this type are valuable for studying aspects of small team behavior, including modeling small adversarial teams. However, they are generally not appropriate for more societal or cultural issues, such as state failure, crowd control, or adaptation in terrorist networks. On the other hand, an ABM could comprise tens of thousands or mil- lions of cognitively simplistic agents doing relatively simple tasks. In this case, interactions among agents would be the result of the agents meeting and greeting each other, trying to occupy the same space, or exchanging or consuming resources. Such models are more common in Âbiology, physics, and the social and organizational sciences; illustrative models are those involving SWARM, REPAST, or MASON. Such models are often used to examine whether the complexity of real-world social processes can arise from agent interactions rather than from the Â complexity of individual agents. Models of this type are valuable for academic research, suggesting possible scenarios, providing very high-level guidance, and studying migra- tion and crowd control. Mid-range models often are comprised of 10 to 10,000 agents with moderately sophisticated learning capabilities. Such models are often Âwritten directly in high-level languages like C++ for reasons of processing speed. In this case, interactions among agents are the result of deliberate decision- making and learning processes that are strongly informed by empirical data. Agent behavior can be quite detailed, such as a detailed mapping of activi- ties taken in a day and the influence of a bioattack on those activities. Such models are increasingly used in such application areas as epidemiology, state failure assessment, crowd control, organizational design, adversarial modeling, and counterterrorism. Models of this type, particularly when
242 BEHAVIORAL MODELING AND SIMULATION they are strongly tied to data and employ socially sophisticated agents, can provide actionable intelligence in the areas listed. Social Sophistication As a general rule, the social sophistication of an ABM Â varies with the number of agents, with the level of sophistication being highest for mid-size populations and lowest for models with only a few agents or with millions of agents. Realistic social behavior requires a certain level of cognitive sophistication (Carley and Newell, 1994). However, many social issues do not emerge as relevant until intermediate-sized social groupings are considered. Typically, models with either a few agents or with millions of agents impose assumptions that limit the use of such models for examining social issues. For example, in ABMs with only a few cognitively sophisticated agents, social factors are typically either ignored or prescribed in terms of a communication and command hierarchy, implying that the structure governing social behavior is time invariant. For example, in models com- prising millions of cognitively simplistic agents, real social networks are typically not modeled. Instead, agents are differentiated using from two to five sociodemographic dimensions and ânetworkâ links are characterized by nearness in a grid. As such, these models are insufficient for modeling terrorist networks. This high level of simplification means that such models rarely generate actionable intelligence. Agents in Grids In many models, particularly those comprising large numbers of cog- nitively simplistic agents, the agents are generally constrained to interact within some form of grid structure. There are two main ways in which agents are laid out in grids: relational and spatial. In a relational approach, each grid cell represents an agent, and the attributes and actions of this agent are determined in part by the attributes and actions of agents in nearby cells. In contrast, in a spatial approach, each grid cell is a location that agents move through (right-left, up-down). Agents consume or leave resources in the cells they occupy, and they interact with the agents they meet in the same or neighboring cells. The classic example of an ABM using a grid is John Conwayâs Game of Life (see Gardner, 1970). Today, many grid-based ABMs are barely more â ee S http://www.econ.iastate.edu/tesfatsi/anetwork for annotated pointers to ABM research on the formation and evolution of social networks.
MESO-LEVEL FORMAL MODELS 243 complex than the original Game of Life, although modern systems use a doughnut-shaped grid (torus) rather than a rectangular grid to avoid edge effects. Grid-based modeling facilitates rapid model development and is supported by agent-based modeling toolkits such as SWARM, REPAST, and Netlogo. However, it is not adequate to support the realistic modeling of four-dimensional social behavior (in space and time) or to capture social network effects in any great depth. In particular, then, ABM frameworks with grid layouts are currently of limited utility for modeling military situations requiring high levels of real- ism. Space-time action sequences and sophisticated social network effects are important factors that need to be carefully accounted for in a variety of military models. For example, they are needed if one is to build a model of adversarial behavior in an urban setting in which an adversary can move from subways to rooftops and can receive shelter from friends. ABM and Learning In ABMs, the agents learn. A major issue is how to model the minds of the cognitive agents who populate ABM frameworks. Should these minds be viewed as logic machines for planning and reasoning with appended data filing cabinets, the traditional artificial intelligence view (Franklin, 1995), or should these minds be viewed as controllers for Â embodied activity in keeping with the artificial life view (Clark, 1997)? On one hand, as with any simulation system, if the purpose of an ABM framework is to determine an optimal design for a fully automated process, there is no particular reason why agent cognition should mimic that of real people. Indeed, this could be positively detrimental to good process perfor- mance. On the other hand, if the purpose is to replicate and forecast human social behavior, then mimicry of real human behavior might be essential to ensure predictive content. As detailed in Brenner (2006), ABM researchers are increasingly mov- ing away from the unconsidered adoption of off-the-shelf machine learning representations, such as conventionally specified genetic algorithms and reinforcement learning algorithms. Some ABM researchers are systemati- cally investigating the performance of alternative learning representations in various multiagent decision contexts. Others are attempting to calibrate their learning representations to empirical decision-making data and human subject experimental data. â ee S http://www.econ.iastate.edu/tesfatsi/aemind.htm for annotated pointers to ABM Âresearch on agent learning representation.
244 BEHAVIORAL MODELING AND SIMULATION ABM and Social Networks Social networks comprise one of the more active research areas in agent-based modeling. One critical issue is the manner in which social networks are determined through deliberative choice of partners as well as by chance and necessity. For example, in economics a key concern has been the emergence of trade networks among collections of buyers and sellers who determine their trade partners adaptively, on the basis of past experi- ences with these partners (Tesfatsion, 1997). A second critical issue concerns the management of a social network for a common (team) goal when participant agents have different motivations for when and how to interact. An example would be the optimal organiza- tion of a corporate enterprise comprising multiple divisions. A third critical issue concerns the disruption of harmful social net- works. For example, research on terrorist networks suggests that they are difficult to destabilize when they have a cellular organization, with partici- pant agents in communication only on a need-to-know or similarity basis. For each of these issues, it is important to consider the extent to which social networks affect the ability to predict social and cultural outcomes with accuracy based on observable structural conditions and institutional arrangements. More precisely, to what extent and with what fidelity does a modeler need to capture social network effects together with structural and institutional effects in order to achieve satisfactory predictive power? For illustration, consider the case of markets. Some types of markets can be expected to display only weak social interaction effects, for example, pool-based wholesale electric power markets under the strong control of a system operator. In this case, the structural aspects of the market (e.g., numbers of buyers and sellers, costs, capacities) and the institutional aspects of the market (e.g., the legal contractual arrangements governing market participation) will presumably be the primary determinants of market out- comes. Other types of markets can be expected to display strong social interaction effects. This is true for labor markets, in which work contracts are highly incomplete and outcomes are strongly dependent on work site interactions between workers and employers. For such a market, given any single structural and institutional starting point, there will presumably be a wide variety of possible outcomes based partly on random social Âinteraction effects. As another example, consider modeling of state failure. A model that examines only the social network among the various stakeholders will not be able to predict state failure with accuracy, nor will a model that â ee http://www.econ.iastate.edu/tesfatsi/anetwork.htm for annotated pointers to ABM S r Â esearch on interaction networks. See, also, the volume of readings edited by Breiger et al. (National Research Council, 2003) and the surveys by Vriend (2006) and Wilhite (2006).
MESO-LEVEL FORMAL MODELS 245 examines only the resources or actions available to the different participant actors. However, by combining these considerations into a single model in which agents are encouraged or discouraged from taking actions by those to whom they are linked, state failure can be better predicted. In summary, applications that require the generation of actionable intel- ligence in social situations will generally require careful consideration of social network effects along with structural and institutional effects. ABM Development Issues A computational laboratory (CL) is a framework that permits the study of complex systems by means of controlled, replicable, computational experiments using an integrated array of specialized software tools. In par- ticular, CLs providing a variety of agent-based tools facilitate the integrated development of ABMs. A number of critical issues arise regarding the development of CLs for ABM applications. For example, should a separate CL be constructed for each application, or should researchers strive for general multifaceted platforms? How can experiÂmental findings be effectively communicated to other researchers by means of descriptive statistics and graphical visualiza- tions without information overload? How might these findings be verified and validated by comparisons with output and data obtained from other sources? How might they tell researchers to look at existing data in different, more dynamic ways? A particularly important unresolved issue is the need to ensure that findings from CL experiments reflect fundamental aspects of a considered problem and not simply the peculiarities of the particular hardware or software platform used to implement the experiments. CLs clearly ease the entry barrier for Âresearchers wishing to use ABM in various problem applications. However, it is important to keep in mind that the use of such integrated development environments permits even novice simulators to build seemingly powerful ABMs in the course of a few months. As a result, we are now seeing thousands of small systems being built by individuals or small teams with little or no training in simulation, and the models are being used to inform critical decision making and policy. In the absence of any accepted criteria for validation (see discussion below and in Chapter 8), it is impossible to judge whether these models are adequate for their intended purposes. On the positive side, the use of ABMs enables the analyst to systemati- cally consider the interaction among more factors and so base decisions on â ee http://www.econ.iastate.edu/tesfatsi/acomplab.htm for annotated pointers to ABM S research on CLs. See also Dibble (2006) for a detailed discussion of CL use for spatial agent- based modeling.
246 BEHAVIORAL MODELING AND SIMULATION a more thorough analysis. On the negative side, the development of ABMs by those not trained in simulation means that the results of the models are often misinterpreted and classic mistakes are often made, which cause the results from the models to reflect incorrect simulation practices rather than interactions among the modeled factors. In summary, great care must be taken in the development of ABM frameworks. Although CLs permit rapid individual development of ABMÂ frameworks, detailed, sophisticated ABM frameworks that produce actionable results often need to be developed by a team working collec- tively for three to five years. It makes sense to use separate teams for data gathering, validation, and usability testing, as each of these areas requires different types of scientific skills. In addition, the team building the model often needs to employ many of the same techniques for development that are used in system engineering. Relevance, Limitations, and Future Directions Military operations intrinsically involve military engagements with rival forces, and forces intrinsically involve equipment and human partici- pants in dynamic motion over geographic terrains. Stated more abstractly, military operations are complex dynamic processes involving multiple, heterogeneous, strategically interacting agents operating through time over spatial landscapes. Framed in this way, the modeling of military operations is seen to be precisely the type of modeling challenge that agent-based modeling is designed to address. What, specifically, are its key advantages for military applications? First and foremost, agent-based modeling provides flexibility. Agents can be modeled as autonomously driven entities operating on their own time scales in fulfillment of individual or group goals. Their methods of operation can be constrained by idiosyncratic personal and cultural con- siderations. They can be equipped with social communication capabilities permitting adaptive information acquisition and transmission. They can survive or not depending on their ability both to secure life-sustaining resources and to manage or prevent life-threatening situations. In particular, agents in ABM virtual worlds can be designed to live in their world with the same degree of flexibility as their real-world counter- parts. System behaviors emerge from the bottom up, through the decen- tralized actions of autonomous agents situated in space and time. This contrasts with a command and control approach to modeling in which outcomes are enforced from the top down. A top-down approach requires that every contingency be anticipated. A bottom-up approach need not anticipate all contingencies, but it must have a sufficiently rich behavioral
MESO-LEVEL FORMAL MODELS 247 repertoire at the individual level so that the system can respond to whatever situation arises. This is precisely the type of modeling flexibility that agent- based modeling provides. Also, agent-based modeling is particularly well suited for studying information diffusion and the evolution of norms, trust, and reputation. The classical game theory approach to these issues seeks to explain behavior on the basis of individual rationality considerations, such as explaining the evolution of norms in terms of anticipations of future reciprocity (Gintis, 2000). In contrast, the ABM approach tends to place equal or greater stress on peer emulation, parental mimicry, and other socialization forces thought to underlie the transmission of culture. ABMs have been used to evaluate the likelihood that the general attitudes of the population would become more pro or con regarding the United States in the face of elections and changes in leadership. ABMs have also been used to forecast state failure, regime change, and the emergence of corruption in various nation-states (Popp et al., 2006). In summary, the issue is not whether agent-based modeling is relevant for modern military operations: it clearly is. The issue is whether it has reached a sufficient stage of development to provide practical support for military operations. Major Limitations ABM frameworks as currently constructed have limitations that could affect their ability to meet critical military needs. This section discusses some of these limitations. Degree of Realism The value of any simulation, including any ABM simulation, is partly tied to the level of realism in the model. Any simulation system is a model and so should be less complex than the real world. However, overÂsimplification results in models so high level or so incorrect that the results can be misinter- preted and so should not be used for policy setting or decision making. The rule of thumb is to make the model only as complicated as it needs to be to address the issue of concern and to achieve the necessary level of fidelity. Adding more rules or equations that increase the realism of the result- ing model should presumably increase its usefulness for decision making. Yet opponents often argue that the more equations or rules, the worse the â ee http://www.econ.iastate.edu/tesfatsi/asocnorm.htm for annotated pointers to ABM S research on the evolution of social norms. See also Young (2006) for a proposed ABM meth- odology for studying the long-run evolution of social norms.
248 BEHAVIORAL MODELING AND SIMULATION model. Arguments include appeals to parsimony, Occamâs razor, under- standability, and so on. A typical argument is that, as the model increases in complexity (number of variables and rules/equations), it becomes increas- ingly likely that the model can be made to fit any possible outcome. (âOver- fittingâ is discussed further in Chapter 9.) This argument derives from econometrics, in which, as the ratio of parameters to data increases, ultimately the data can be completely and per- fectly modeled. This argument, however, is not directly applicable to ABMs. In them the addition of new rules and equations serves to increase the number of outcomes or dependent variables (data) that can be generated; the data are not given a priori as in econometrics. Moreover, the addition of empirically based rules and equations can increase the plausibility of these generated outcomes by reducing the possibility of implausible results. The realism of ABMs can be increased, and their military value increased, as they are linked to real data. Most groups that build them have contrasted, at best, the results of one dependent variable with real data. Only a few agent-based models, such as some recently created for the Defense Advanced Research Projects Agency or the BioWar system, use massive amounts of real data to set the input specifications of the models and other data to validate the system. In general, this requires the linking of the models to database systems. The key technical challenge here is that, as the ontology in the database changes, the model needs to be augmented. There are currently no tools to facilitate such changes. A second challenge is that, for validation, it is important to have the model produce data in the same form as the real data, that is, to create a comparable database. There are currently no standardized tools for doing statistical comparison of data in two identically structured databases. Model Trade-Offs ABMs using cognitively sophisticated agents tend to require the use of knowledge engineering techniques. Such models tend to be special purpose and permit minimal reuse. The key value of such models is to take the place of human teams in war-gaming situations, equipment testing, and design situations and to evaluate processes that facilitate team behavior. In general, these models use various cognitive architectures with multiagent components added and so are often limited to only a small number of agents. Their strength is looking at detailed task-related behavior. As previously noted, such models tend to use predefined social interactions. This limits their use in war games because the ABMs do not exhibit a full range of adaptive i Ânteraction but model only limited task-based communications and actions. The typical grid-based ABMs, with millions of cognitively unsophisti- cated agents, are generally useful only for high-level explorations of gen-
MESO-LEVEL FORMAL MODELS 249 eral concepts. They are valuable for starting groups to think outside the box and for provoking discussions. These models are rarely sophisticated enough to be used as an adaptive adversary in war-gaming or for evaluating task-based behavior. The strength of these models is their Â ability to look at population-level trends resulting from local action. As such, they show promise in such areas as marketing, impact of psychological operations, information diffusion studies, and disease transmission studies. Rarely do such models generate actionable intelligence. Now consider dynamic network ABMs tied to empirical data. Such models utilize agents with moderate levels of cognitive sophistication and high levels of social sophistication. This results in models that can be used for war-gaming to look at adaptive adversaries. Given current technology, this combination results in models that can handle more agents but that run more slowly. The strength of these models lies in representing and reasoning about fairly large-scale units, such as the armyâs unit of action, cities at 20 percent population, or terrorist networks. The added cognitive and social sophistication inherent in these models makes it possible to produce action- able results. However, getting a model to the point of producing actionable results takes a multiperson, multiyear data collection effort on top of a multiyear model development effort. Modeling of Actions One of the key factors limiting ABMs from a military perspective is the modeling of actions. Currently actions can be modeled at a very high level (pro-con, hostile, friendly, or neutral) or at a very detailed level (fire a par- ticular weapon). There is neither a middle ground nor a hierarchy relating actions at one level to another. Therefore, efforts to model actions tend to be either very generic or single use. A basic ontology of actions is needed for the state of the art to advance. Research and Development Requirements Several requirements can be identified for the further development of ABMs that could be of use in military settings. The next section discusses tool development, data farming, linkages of agent-based modeling to other modeling efforts, and the development of the human resources and exper- tise needed to support ABM development. Tool Development Key advances and applicability to military modeling require agent- based modeling and network analysis techniques to be integrated into tool
250 BEHAVIORAL MODELING AND SIMULATION chains. For example, pattern discovery techniques can be used to derive equations from historical data that can then be used in ABMs to evolve future systems. ABM techniques can be used to evaluate courses of action and to suggest areas for further data collection. Combining these techniques will enable new types of problems to be solved; for example, combining social network metrics with pattern discovery techniques is the key to build- ing an understanding of how networks grow and evolve. This is not to suggest that the military should move to large integrated behavioral modelsâquite the contrary. What is needed is increased interÂ operability of the tools. The development of ABM CLs and the explosion of network analytic tools are putting social behavioral modeling into the hands of the masses. Moreover, these trends are leading to the development of many small, single-purpose tools. This should be taken advantage of by encouraging interoperability (this is also discussed further in Chapter 8). It is important to note that it would not be feasible to require all tools to be written in a single language or to require the use of a single frame- work; rather, the solution needs to enable the integration of models not only from diverse domains but also in diverse languages. Multiple models, visualization tools, and the like should be available to address diverse prob- lems, but in such a way that data (real and virtual) can be shared easily among the various tools. There are a variety of things needed to support such interoperability. Standards for the interchange of relational data need to be developed. Behavioral modeling tools need to be web enabled, and XML input-output (IO) languages need to be developed. A uniform vocabulary for describ- ing relational data also needs to be developed; this is particularly critical because the tools and metrics are coming out of at least 20 different scientific fields. For defense and intelligence applications, common platforms and data sharing standards need to be explored and developed so that tools written in the unclassified realm can be rapidly moved, without complete redesign, to the classified realm. Enabling interoperability and providing a platform and common ontologies for these tools will enable novel problems to be more rapidly addressed by regrouping existing models. It will also enable various subject matter experts to interact through the interaction of their models. In turn, this will enable a broader approach to problems, reduce the likelihood of biased solutions, and facilitate rapid development and deployment. â hese T fields include anthropology, sociology, psychology, organization science, marketing, physics, electrical engineering, geology, ecology, economics, biology, bioinformatics, health services, forensics, artificial intelligence, robotics, computer science, mathematics, statistics, information systems, medicine, civil engineering, communication, and rhetoric.
MESO-LEVEL FORMAL MODELS 251 Current tools are either very data-greedy or become more valuable as they are linked to real data. However, there is a dearth of relevant data currently available in clean preprocessed form. Thus, to reduce the time analysts spend on data collection and to increase the time they spend on analysis, automated and semiautomated tools for data gathering, cleaning, and sharing are needed. Such tools should include natural language process- ing tools for extracting relational data from audio and text sources, âweb- scrapingâ tools, automatic ontology generators, and visual interpretation tools to extract network data from photographs and visual images. Appropriate subtools for node identification, entity extraction, thesau- rus creation, and other functions are also needed. The development and availability of these tools in an interoperable environment are critical for providing masses of data that can be used for model tuning and validation. Moreover, these tools reduce time spent on data collection and thereby free the analystsâ time for analysis. More rapid data collection would also mean the availability of more datasets for doing meta-analyses, thereby enabling improvements in the theoretical foundations of the field and in the understanding of social behaviors. Finally, these tools are essential for providing the wealth of data needed by social behavioral models to make reasonable forecasts or to provide reasonably accurate analyses of situa- tions and organizations. Improved speed for many of the algorithms could be provided by computer architectures designed for relational data or by the use of special integrated circuits with embedded versions of the less scalable algorithms. Note this would enable a speed savings beyond that afforded by the use of current vector technology. Such technology would facilitate faster pro- cessing and enable more real-time solutions, particularly for large-scale networks. To reduce the âartâ aspect of interpretation in this field, a living archive of collected network data is needed, replete with information on metrics for the nodes in each dataset. Such an archive could be used to set context information. For example, such information could be used to evaluate whether the density of particular networks is exceptionally high or low or to identify exceptional values of connectedness of individuals. Such an archive would facilitate meta-analysis and comparative analysis. This is critical for improving the theoretical foundations of the field as well as for the understanding of social behavior. Forecasting and Possibility Analysis Of the models described here, those that have shown the most promise in terms of forecasting are the voting models, the dynamic network Âmodels (that combine agent-based technology and meta-matrix of relations), and
252 BEHAVIORAL MODELING AND SIMULATION the social influence models. These models have had limited success in forecasting voting outcomes, changes in beliefs and attitudes at the macro level, and identifying emergent new leaders. For other modeling techniques, including the ABM and system dynamic techniques for complexity model- ing, the models are best at providing insight into the space of possibilities, that is, demonstrating what possible futures might exist and their relative likelihood. However, for these models to provide an adequate map of the possibilities (a reasonable response surface), the models need to be run a vast number of times under diverse scenarios; hence, as is discussed in the next section, there is a need for placing these models in a data farming environment. One question that arises is: How can these models be made more pre- dictive? This topic, in and of itself, is quite complex and a full treatment is beyond the scope of this study. However, several factors are worth noting. As more of these models are placed in data farming environments, statistical tools are developed for mining the vast data so generated, and repositories of meta-matrices are developed and shared with scientists for testing and validating, one can expect that many of these models will become more reliable in their forecasts. However, there will still be many classes of social phenomena for which prediction, of the form used in engineering and physics, will simply not be possible due to the lack of stationarity in the underlying social processes, the paucity of data, and the lack of continuity in key variables. A second question often arises regarding the concern that, if the models are truly predictive, the mere act of making a prediction public will cause actors to change their behaviors and so alter the outcome. While this issue is addressed in other sections of this report, several key factors directly related to the nature of the models described here are worth mentioning. For most of the models described here, other than the simple voting Âmodels, making the models transparent to the public (so that others can infer the predictions) or making the predictions themselves public is not likely to invalidate the predictions. There are three basic reasons for this: lack of temporal forecasting, level of specificity, and hyper-confluence. Temporal forecasting tends to be weak and predictions are often vague in terms of when something will occur; rather than point predictions, most predictions are of the form âA will likely occur after Bâ or âat some time in the future more than two weeks but less than two years from now.â Most models produce rather general results, such as that a state will fail, civil violence is likely to erupt, or corruption will increase, rather than the more specific âthe state will fail due to a regime change where General X takes overâ or âcivil violence will take the form of riots in these five citiesâ or âcorruption will increase the most in the area of infrastructure development in county X.â Finally, most models generate a prediction due to hyper-Âconfluence,
MESO-LEVEL FORMAL MODELS 253 that is, the strongest predictions are those for which there are a large num- ber of interconnected causes that weave together in complex ways. But single actors can best counter a specific event that is likely to occur at a specific time with only one or two actions or activities. Even with sufficient research funding, improved theory, and available data to overcome the issues of vague temporal forecasting and lack of specificity; the problem of hyper-confluence will remain. That is one of the key reasons why social and behavioral models need to be driven by the science of the possible, rather than the traditional science of point predictions involved in traditional physical science and engineering models. Data Farming ABMs designed for applied settings need to be placed in data farming environments. These environments need to be augmented with special- purpose tools for running massive virtual experiments. These tools should enable improved visualization and analysis and facilitate the development of semiautomated response surface generators. Current data farming tools often are cumbersome to use, require code modification of the ABM, and are limited by the processor speed and storage capabilities of the machines that they run on. In order for ABM frameworks to run routinely in data farming environ- ments, more flexible environments need to be developed and made easily available to researchers. Moreover, ABM frameworks need to be developed with wrappers, so that they can be placed in these environments. Standard- ized IO formats need to be developed. By routinely placing ABM frame- works in a data farming environment, a better understanding of the space of possibilities predicted by the frameworks will be derived. This will enable ABM frameworks to better support policy and decision making. Currently, when ABM frameworks are used to inform policy and criti- cal decisions, they are typically run only a few times in carefully controlled computational experiments. While this approach enables the analyst to explore more possibilities more systematically than not using a simulation, it still leaves open the possibility that errors might be made if the results are generalized beyond the scope of the experiment. By placing ABM frame- works in a data farming environment, the number of computational experi- ments conducted, the space of possibilities examined, and the scope of â A â wrapper is a software layer used to change the interface of a component or to give new properties, such as fault tolerance or security, to the interaction between components. Software wrappers are often used to glue existing subsystems into a larger system with new properties and functions. The wrappers know the protocols needed to make the subsystems work together, even if they were not originally designed for a common purposeâ (Webber, 1997, p. 1).
254 BEHAVIORAL MODELING AND SIMULATION analyzed conditions can be expanded, often by several orders of magnitude, thus providing a stronger basis for decision making. Furthermore, once an ABM framework has been validated, the response surface equivalent can be used as a ârapidâ model in training situations in which the users do not have time to wait for an ABM experiment to finish running. Cross-Disciplinary Initiatives Another avenue that may promote major breakthroughs is the linkage of ABM social behavioral modeling to gaming environments, particularly online multiplayer games such as Everquest and Americaâs Army (see Chap- ter 7). Research initiatives that explore the link of ABM social behavioral modeling to gaming tools may be valuable. Possible research areas include using agent-based modeling to explore the realism of the social behaviors exhibited in gaming models; using it to provide flexible opponents or to make the apparent number of game players larger and so force players to think about group scale issues; and using agent-based modeling to track and analyze game behaviors using dynamic network analysis techniques. Key benefits here would be improved training tools and visual what-if scenario evaluation. As previously noted, additional ABM development needs to be done in a number of areas. These include attachment of ABM frameworks to data streams, improved ABM visualization, metric ABM robustness studies, and so on. Moving ahead in these areas will require linking social networks to other types of data, such as location and event information, and linking diffusion theory to other forms of theory, such as action and cultural theory. This will require the funding of both basic and applied research. It will also require an increased recognition for, and acceptance of, applied social sci- ence research in universities. Currently there are a number of funded research efforts in the areas of cultural modeling, geospatial link analysis, and adversarial modeling, all of which are supporting work along these lines. A key to much of this work is that it combines dynamic network analysis with geospatial rea- soning or anthropological data-gathering techniques. Much of this work is applied, directed at providing usable systems in several years. This is a positive development, particularly when such modeling efforts are based on strong empirical and theoretical foundations. However, there is still a huge amount of basic research to be done in such areas as the development of an ontology for tasks, a unified model of culture, or even a shared definition of culture. Relatively little research funding is being directed to the basic research questions in this area. The key here is not simply to invest in the social sciences but to invest in the mathematical and computational social sciences to engender the
MESO-LEVEL FORMAL MODELS 255 development of work that will support defense needs. The benefit will be an improved understanding of basic social and cultural phenomena. Another benefit will be a decrease in the development of misleading models that appear to be social but that are not theoretically or empirically sound. At the same time, most of the research community, particularly in the social sciences, is not focusing on strongly applied problems. The mere idea of hard deliverables, while accepted as common practice in engineering and computer science, is contrary to the basic culture of most social sci- ence departments. Thus, while there is a strong need for quantitative social science modeling on defense issues, there is a dearth of social scientists involved in and trained to do applied work. Building Expertise The lack of highly trained professionals is a key difficulty in this area. Universities need to expand their undergraduate social science curricula to include more of the mathematical and computational social sciences. In particular, undergraduate courses should be routinely taught that cover SNA and agent-based modeling, and that permit the mastery of ABM programming tools. Universities need to encourage and facilitate applied research. New curricula are needed that have an engineering style but that are focused on social and policy applications. Masterâs programs that combine social and computational science need to be developed. Military universities, such as West Point and the Naval Postgraduate School, should also offer social network courses and possibly ABM courses, particularly those for evolving networks, and they should integrate dynamic network measures of shared situation awareness, leadership, and power into the standard curriculum. The development of these curricula and degree programs is vital to the nationâs intellectual strength in order to remain at the forefront in this area. The clear benefit of these programs will be a stronger workforce of computational social analysts capable of developing and using social behavioral models. Analysts engaging in ABM but trained in computer science, engineer- ing, or physics should work in teams with social scientists to avoid duplicat- ing work already done or making commonsense assumptions about social processes that have no empirical bases. Corporations need to provide time and resources for selected personnel to become jointly trained in computer and social science, either by increasing the number of personnel sent to masterâs programs, bringing in relevant faculty to teach short courses, or engaging in more joint research with universities as equal partners (in which the university provides the missing skill, social or computational). The key advantage of teaming is that it will enable improved model development
256 BEHAVIORAL MODELING AND SIMULATION and will serve as a stop-gap until more computational social analysts are trained. Expected Outcomes Across the board, success in the activities outlined above would facili- tate the rapid development and deployment of agent-based modeling. The advantage is that it enables systematic reasoning about various courses of action in a wide range of complex environments. More courses of action could be evaluated in less time and more systematically than is done with conventional table-top war-gaming or current non-computer-assisted analy- sis of relational data. The dynamic social network and ABM tools outlined above reduce the time spent on data processing and increase time spent on analysis and interpretation. They would facilitate what-if analysis and could ultimately support near-real-time what-if analysis in the field. This would be a valuable force multiplier. In summary, the activities listed above would increase the maturity of the modeling field, improve scientific theory, facilitate rapid linking of computational models to empirical data, particularly network data in a unified reasoning framework to solve novel problems, and encourage new discoveries. These activities would also promote the development of a new science that combines computation and society, just as the previous join- ing of computer science, design, and psychology led to the new science of human-computer interaction. References Arrow, K.J. (1951). Social choice and individual values. Hoboken, NJ: John Wiley & Sons. Arthur, W.B. (2006). Out-of equilibrium economics and agent-based modeling. In L. ÂTesfatsion and K.L. Judd (Eds.), Handbook of computational economics, volume 2: Agent-based computational economics. Amsterdam, The Netherlands: Holland/Elsevier. Axelrod, R. (1997). Advancing the art of simulation in the social sciences. In R. Conte, R. Hegselmann, and P. Terna (Eds.), Simulating social phenomena (pp. 21â40). Berlin: Springer. Borgatti, S.P., and Everett, M.G. (1992). Notions of position in social network analysis. Â ociological Methodology, 22, 1â35. S Borgatti, S.P., and Everett, M.G. (2006). A graph-theoretic framework for classifying centrality measures. Social Networks, 28(4), 466â484. Borgatti, S.P., and Foster, P. (2003). The network paradigm in organizational research: A review and typology. Journal of Management, 29(6), 991â1013. Borgatti, S.P., Everett, M.G., and Shirey, P. (1990). LS sets, lambda sets and other cohesive subsets. Social Networks, 12(4), 337â357. Breiman, L. (2001). Statistical modelingâThe two cultures. Statistical Science, 16, 199â231. Brenner, T. (2006). Agent-learning representation: Advice on modeling economic learning. In L. Tesfatsion and K.L. Judd (Eds.), Handbook of computational economics, volume 2: Agent-based computational economics. Amsterdam, The Netherlands: Holland/Elsevier.
MESO-LEVEL FORMAL MODELS 257 Butts, C.T., and Carley, K.M. (in press). Structural change and homeostasis in organizations: A decision-theoretic approach. Journal of Mathematical Sociology. Carley, K.M. (2003). Dynamic network analysis. In National Research Council, Dynamic social network modeling and analysis: Workshop summary and papers (pp. 133â145). R. Breiger, K.M. Carley, and P. Pattison (Eds.), Committee on Human Factors. Board on Behavioral, Cognitive, and Sensory Sciences, Division of Behavioral and Social Sciences and Education. Washington, DC: The National Academies Press. Carley, K.M. (2004). Estimating vulnerabilities in large covert networks using multilevel data. In Proceedings of the NAACSOS 2004 Conference, Pittsburgh, PA. Available: http:// www.casos.cs.cmu.edu/events/conferences/2004/2004_proceedings/Carley,Kathleen.doc [accessed Feb. 2008]. Carley, K.M., and Newell, A. (1994). The nature of the social agent. Journal of Mathematical Sociology, 19(4), 221â262. Carley, K.M., Columbus, D., DeReno, M., Reminga, J., and Moon, I.-C. (2007). ORA userâs guide. (Report No. CMU-ISRI-07-115). Carnegie Mellon University School of Computer Science, Institute for Software Research. Chen, H., and Lynch, K.J. (1992). Automatic construction of networks of concepts character- izing document databases. IEEE Transactions on Systems, Man, and Cybernetics, 22(5), 885â902. Clark, A. (1997). Being there: Putting brain, body, and world together again. Cambridge, MA: MIT Press. Converse, P.E. (1964). The nature of belief systems in mass publics. In D.E. Apter (Ed.), Â deology and discontent (pp. 206â261). London, England: Free Press of Glencoe. I Davis, A., Gardner, B.B., and Gardner, M.R. (1941). Deep south: A social anthropological study of caste and class. Chicago, IL: University of Chicago Press. de Toqueville, A. (1835). Democracy in America. London, England: Saunders and Otley. Dibble, C. (2006). Computational laboratories for spatial agent-based models. In L. ÂTesfatsion, and K.L. Judd (Eds.), Handbook of computational economics, volume 2: Agent-based computational economics. Amsterdam, The Netherlands: Holland/Elsevier. Festinger, L. (1957). A theory of cognitive dissonance. Palo Alto, CA: Stanford University Press. Franklin, S. (1995). Artificial minds. Cambridge, MA: MIT Press. Freeman, L. (2004). The development of social network analysis: A study in the sociology of science. Vancouver, British Columbia: Empirical Press. Freeman, L.C., White, D.R., and Romney, A.K. (Eds.). (1991). Research methods in social network analysis (revised edition). Piscataway, NJ: Transaction Press. Friedman, J. (2005). Popper, Weber, and Hayek: The epistemology and politics of ignorance. Critical Review, 17(1â2). Available: http://www.criticalreview.com/2004/pdfs/ignorance_ article.pdf [accessed March 2008]. Gardner, M. (1970). The fantastic combinations of John Conwayâs new solitaire game âLife.â Scientific American, 223, 120â123. Getoor, L., and C. Diehl (2005). Link mining: A survey. SIGKDD Explorations, 7(2). Avail- able: http://www.cs.umd.edu/~getoor/Publications/getoor-kddexp05.pdf [accessed April 2008]. Getoor, L., Friedman, N., Koller, D., and Taskar, B. (2001). Probabilistic models of relational structure. In Proceedings of the International Conference on Machine Learning. Avail- able: http://www.cs.umd.edu/~getoor/Publications/jmlr02.pdf [accessed April 2008]. Getoor, L., Friedman, N., Koller, D., and Taskar, B. (2002). Learning probabilistic models of link structure. Journal of Machine Learning Research, 3, 679â707. Available: http:// www.jmlr.org/papers/volume3/getoor02a/getoor02a.pdf [accessed Feb. 2008].
258 BEHAVIORAL MODELING AND SIMULATION Gibbard, A. (1973). Manipulation of voting schemes: A general result. Econometrica, 41(4), 587â601. Gintis, H. (2000). Game theory evolving: A problem-centered introduction to modeling Â trategic interaction. Princeton, NJ: Princeton University Press. s Gladwell, M. (2000). The tipping point: How little things can make a big difference. Boston, MA: Little, Brown, and Company. Goldberg, H.G., and Senator, T.E. (1998). Restructuring databases for knowledge discovery by consolidation and link formation. In Proceedings of the 1st International Conference on Knowledge Discovery and Data Mining, Quebec, Montreal. Available: http://citeseer. ist.psu.edu/cache/papers/cs/3649/http:zSzzSzeksl-www.cs.umass.eduzSzailazSzgoldberg- senator.pdf/goldberg95restructuring.pdf [accessed Feb. 2008]. Goldberg, H.G., and Wong, R.W.H. (1998). Restructuring transactional data from link analysis in the FinCEN AI system. In Proceedings of the 1998 AAAI Fall Symposium on Artificial Intelligence and Link Analysis. Available: http://kdl.cs.umass.edu/events/ aila1998/Âgoldberg-wong.pdf [accessed April 2008]. Green, D., and Shapiro, I. (1994). Pathologies of rational choice theory. New Haven, CT: Yale University Press. Harvey, J.B. (1974). The Abilene paradox and other meditations on management. Organiza- tional Dynamics, 3(1). Hauck, R.V., Atabakhsh, H., Ongvasith, P., Gupta, H., and Chen, H. (2002). COPLINK concept space: An application for criminal intelligence analysis. IEEE Computer Digital Government Special Issue, 35(3), 30â37. Heider, F. (1988). The notebooks: Balance theory, volume 4. New York: Springer Verlag. Jensen, D., and Neville, J. (2002). Linkage and autocorrelation cause feature selection bias in relational learning. In Proceedings of the Nineteenth International Conference on Â achine Learning (pp. 259â266), Sydney, Australia. Available: http://citeseer.ist.psu.edu/ M cache/papers/cs/30391/http:zSzzSzkdl.cs.umass.eduzSzpaperszSzjensen-neville-icml2002. pdf/jensen02linkage.pdf [accessed Feb. 2008]. Kollman, K., and Page, S.E. (2006). ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ï¿½ Computational methods and models of politics. In L. T Â esfatsion and K.L. Judd (Eds.), Handbook of computational economics, volume 2: Agent-based computational economics. Amsterdam, The Netherlands: Holland/Elsevier. Kubica, J., Moore, A., Schneider, J., and Yang, Y. (2002). Stochastic link and group detec- tion. In Proceedings of the Eighteenth National Conference on Artificial Intelligence (pp. 798â804), Menlo Park, CA: AAAI Press/MIT Press. Available: http://www.cs.cmu. edu/~schneide/AAAI02_GDA.pdf [accessed April 2008]. Kubica, J., Moore, A., and Schneider, J. (2003). Tractable group detection on large link data- sets. In X. Wu, A. Tuzhilin, and J. Shavlik (Eds.), The Third IEEE International Confer- ence on Data Mining (pp. 573â576). Washington, DC: IEEE Computer Society. LatanÃ©, B. (1996). Dynamic social impact: The creation of culture by communication. Journal of Communication, 46, 13â25. Lee, R. (1998). Automatic information extraction from documents: A tool for intelligence and law enforcement analysts. In Proceedings of the AAAI Fall Symposium on Artificial Intel- ligence and Link Analysis (pp. 63â67). Available: http://citeseer.ist.psu.edu/cache/Âpapers/ cs/15373/http:zSzzSzeksl-www.cs.umass.eduzSzailazSzlee.pdf/richard98automatic.pdf [accessed Feb. 2008]. Marchant, T. (2000). Does the Borda rule provide more than a ranking? Social Choice and Welfare, 17(3), 381â391. Milo, R., Shen-Orr, S., Itzkovitz, S., Kashtan, N., Chklovskii, D., and Alon, U. (2002). Network motifs: Simple building blocks of complex networks. Science, 298(5594), 824â827.
MESO-LEVEL FORMAL MODELS 259 Moon, I.-C., and Carley, K.M. (2007). Modeling and simulation of terrorist networks in social and geospatial dimensions. IEEE Intelligent Systems, Special Issue on Social Computing, 22(5), 40â49. National Research Council. (2003). Dynamic social network modeling and analysis: Workshop summary and papers. R. Breiger, K. Carley, and P. Pattison (Eds.), Committee on Human Factors. Board on Behavioral, Cognitive, and Sensory Sciences, Division of Behavioral and Social Sciences and Education. Washington, DC: The National Academies Press. Nemeth, C.J. (1986). Differential contributions of majority and minority influence. Psycho- logical Review, 9(1), 23â32. Page, S.E. (2007). The difference: How the power of diversity creates better groups, firms, schools, and societies. Princeton, NJ: Princeton University Press. Page, S.E. (2008). Agent-based models. In L. Blume and S. Durlauf (Eds.), The new Palgrave dictionary of economics, second edition. Hampshire, England: Palgrave Macmillan Ltd. Popp, R., Kaisler, S.H., Allen, D., Cioffi-Revilla, C., Carley, K.M., Azam, M., Russell, A., Choucri, N., and Kugler, J. (2006). Assessing nation-state instability and failure. Paper presented at the Aerospace Conference IEEE 2006, March, Big Sky, MT. Available: http://ieeexplore.ieee.org/iel5/11012/34697/01656054.pdf?tp=&isnumber=&arnumber =1656054 [accessed April 2008]. Saari, D.G. (2001). Decisions and elections: Explaining the unexpected. Cambridge, England: Cambridge University Press. Saari, D.G. (2006). Which is better, the Condorcet or Borda winner? Social Choice and Â elfare, 26(1), 107. W Satterthwaite, M. (1975). Strategy-proofness and Arrowâs conditions: Existence and cor- respondence theorems for voting procedures and social welfare functions. Journal of Economic Theory, 10, 187â217. Schelling, T.C. (1978). Micromotives and macrobehavior. New York: W. W. Norton. Snijders, T. (2001). The statistical evaluation of social network dynamics. In M. Sobel and M. Becker (Eds.), Social methodology dynamics (pp. 361â395). Boston and London: Basil Blackwell. Taskar, B., Abbeel, P., and Koller, D. (2002). Discriminative probabilistics models for rela- tional data. Paper presented at the 18th International Conference on Uncertainty in Artificial Intelligence. Available: http://www.biostat.wisc.edu/~page/rmn.pdf [accessed Feb. 2008]. Taskar, B., Wong, M.F., and Koller, D. (2003). Learning on the test data: Leveraging Â unseenâ features. Paper presented at the 20th International Conference on Machine â Learning, August, Washington, DC. Available: http://ai.stanford.edu/~koller/Papers/ Taskar+al:ICML03.pdf [accessed Feb. 2008]. Tesfatsion, L. (1997). A trade network game with endogenous partner selection. In H. Amman, B. Rustem, and A.B. Whinston (Eds.), Computational approaches to economic problems (pp. 249â269). Dordrecht, The Netherlands: Kluwer Academic. Tsvetovat, M., Reminga, J., and Carley, K.M. (2004). DYNETML: Interchange format for rich social network data. (CASOS Technical Report #CMU-ISRI-04-105). Pittsburgh, PA: Carnegie Mellon University, School of Computer Science, Institute for Software Research International. Vriend, N. (2006). ACE models of edogenous interactions. In L. Tesfatsion and K.L. Judd (Eds.), Handbook of computational economics, volume 2: Agent-based computational economics. Amsterdam, The Netherlands: Holland/Elsevier. Wasserman, S., and Faust, K. (1994). Social network analysis: Methods and applications. New York: Cambridge University Press.
260 BEHAVIORAL MODELING AND SIMULATION Webber, F. (1997). Software wrappers to support nonstop computing. Paper presented at the 20th National Information Systems Security Conference, October, Baltimore, MD. Avail- able: http://csrc.nist.gov/nissc/1997/proceedings/730.pdf [accessed Feb. 2008]. Wilhite, A.W. (2006). Economic activity on fixed networks. In L. Tesfatsion and K.L. Judd (Eds.), Handbook of computational economics, volume 2: Agent-based computational economics. Amsterdam, The Netherlands: Holland/Elsevier. Young, H.P. (2006). Social dynamics: Theory and applications. In L. Tesfatsion and K.L. Judd (Eds.), Handbook of computational economics, volume 2: Agent-based computational economics. Amsterdam, The Netherlands: Holland/Elsevier. Zeggelink, E. (1994). Dynamics of structure: An individual oriented approach. Social Net- works, 16(4), 295â333.