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.
Part II STATE OF THE ART IN ORGANIZATIONAL MODELING 89
Part II State of the Art in Organizational Modeling P art II reviews the multitude of individual, organizational, and societal (IOS) modeling approaches, methods, and tools that are potentially useful for addressing the military modeling needs described in Chap- ter 2. Models take many forms, ranging from loose conceptual models to precise mathematical models (Lave and March, 1975). They include agent- based models, cognitive models, expert systems, dynamical systems, and input-output models. Here we survey and explore many different types of models relevant to our questions. We describe each, show their strengths and limitations, and discuss research and development efforts that could make the approaches more useful for addressing military modeling needs. The diverse expertise of the committee members contributed greatly to the completeness of this review but also made it challenging to agree on an organizing framework for presenting the review results. Refined through multiple iterations, the organizing framework that we developed represents a significant product of the study. Categories of Models: Initial Empirical Results As a first step in organizing our review, we took an empirical approach to organizing the various terms and approaches used in IOS modeling. Using the methods of cultural domain analysis (see Chapter 3 for a descrip- tion), we developed a perceptual map of the field of modeling based on committee membersâ perceptions. 91
92 BEHAVIORAL MODELING AND SIMULATION Methodology The first step in our investigation was to collect âfree listsâ of models from each member of the committee. Effectively, we asked, âWhat are all the kinds of models you can think of?â A large number of unique âkinds of modelsâ were elicited with little overlap, implying that the domain itself lacks a high degree of cultural coherence. A total of 240 items were elicited, with approximately 35 items per member. Much of this lack of overlap was due to differences in the level of specificity for the kinds of models listed. For example, some of the items were specified at the level of named models, such as DyNet, EpiSims, NetWatch, etc., while others were at a very general level, such as conceptual models or verbal models. Aggregating across all lists, a master list of distinct terms was obtained after standardizing word forms. Also, an attempt was made to keep all items at the same level of specificity, in this case at a more general level. The second step was to take the 38 most frequently mentioned items at a more general level of specificity and construct a pile-sorting task, in which each committee member was asked to sort the items into piles according to how similar the kinds of models are. They could use as many or as few piles as they wished. The task was conducted online using interview software that simulated cards and allowed the virtual cards to be placed into piles. When they were done, the program recorded the membership of each pile. Then, an aggregate proximity matrix X, whose rows and columns corre- sponded to âkinds of models,â was constructed such that each cell Xij of the matrix recorded the number of respondents that placed the ith kind of model in the same pile as the jth kind of model. The final step was to visualize this proximity matrix using a standard network visualization package called Netdraw (Borgatti, 2002). In this approach to visualization, a line is drawn to connect two items if the simi- larity of the two items exceeds a certain user-defined threshold (DeJordy, Borgatti, Roussin, and Halgin, 2007; Johnson and Griffith, 1998). Results The resulting map is shown in Figure II-1. In the map, a line is drawn between two modeling techniques if at least 28 percent of the respondents placed the items together in the same pile. (A cutoff of 28 percent was chosen because above that level the main section of the network becomes disconnected.) The results show three basic clusters of modeling techniques. The first cluster, at the top left of the map, consists of multiagent models in which the agents are connected to each other by social ties or interactions. In these models, the combination of agents and ties forms a single interconnected
Multiagent Representative_agent Simulation Cellular_automata Multiagent_network Agent-based Social_network Computational Dynamic_network Process System_Dynamics Dynamical_systems Mathematical Differential_equation Markov Difference_equation Linear Time_series Multiattribute Nonlinear Statistical Cultural Equilibrium Group_decision-making Organizational_culture Econometric Organizational_learning Input-Output Cognitive Game_theory Optimization Machine_Learning Genetic Influence Expert_systems Risk Behavioral Conceptual Decision_theory 93 FIGURE II-1â Perceived similarities among types of models. Part II-1.eps broadside
94 BEHAVIORAL MODELING AND SIMULATION and interdependent system. For convenience, we refer to the models in this cluster as the computational network models cluster (although it contains some models for which this would not be the ideal name). The computational network models cluster is connected to the next cluster via the system dynamics node. This new cluster consists of low-level statistical and mathematical techniques that have broad application across many different settings. Although these techniques are often thought of as tools rather than models, statisticians would recognize that they do indeed constitute models. For convenience, we refer to this cluster as the math- ematical systems models cluster. At the bottom right of Figure II-1, there is another cluster of models focused on the cognition or culture of the agents. We call these the cogni- tive models. The difference between these and the computational network models at the top of the map is one of emphasis rather than substance. The cognitive models are defined by their focus on the details of cognition. The objective of the cognitive models is to understand the patterns of who believes or chooses what. In contrast, the computational network models are defined by the processes that the modeler builds into the system and may not represent cognition at a detailed level. The outcomes of the com- putational network models may well be the same as those of the cognitive models, and the processes of the cognitive models often involve the same multiagent interactions of the computational network models: it is only the focus of the investigation that is different. Finally, as noted at the bottom left, three model typesâinfluence, behavioral, and conceptualâdid not cluster with other types. Four-Part Organizing Framework for Models On the basis of the empirical clustering results described above and further discussion, the committee developed a four-part categorization for reviewing modeling approaches: (1) macro models, (2) micro models, (3) meso models, and (4) integrated, linked micro-meso-macro models. No single one of these approaches is the correct one, and the best model- ing approach depends on the nature of the problem to be solved. It is a common theme throughout this book that models constitute âuse-driven researchâ (Stokes, 1997) and cannot be developed or evaluated without an in-depth understanding of the uses to which they are to be put. A macro model considers interactions between macro-level variables, such as unemployment, crime, education, poverty, and resources. Macro modeling approaches like system dynamics enable one to identify feedbacks and to see system-level effects without getting bogged down in details. At the other extreme, one can model the cognitive or affective processes of individual actors or at least their outcomesâindividual decisions and
STATE OF THE ART IN ORGANIZATIONAL MODELING 95 actions. These more micro modeling approaches include cognitive models from psychology, expert systems models, and rational choice models, which include game theory and decision theory. Fifty years ago, this distinction between micro and macro would have been thought sufficient. One can look at the trees, or one can look at the forest. Over time, social scientists have come to appreciate the importance of the level in betweenâthe meso level (Miller and Page, 2007). To com- plete the metaphor, one can think of stands of aspen trees with a shared root system. The stand is a part of the forest, yet it does not function merely as the sum of its individual trees, given the sharing of resources. The social analog of a shared root system is social capital. People join movements, participate in riots, and support government in part based on the actions of their friends and peers. Predictions based on individual attributes can almost always be improved by adding in social factors. We highlight two types of meso models: network models and agent- based models. Both modeling approaches have produced flurries of atten- tion over the past decade. Network models allow one to formalize, measure, and test loose conceptions of social capital, centrality, and connectedness. Agent-based models allow one to include diverse, purposive agents who interact in space and time. As the name suggests, agent-based models origi- nate with the agents, but these agents can self-organize, creating emergent meso-level structures that take on meaning and have predictive value. The fourth category links micro, meso, and macro models. Agent-based models, and to some extent game theory and network models, achieve this double linkage. Yet only recently have researchers begun to create hybrid models that include agents who employ sophisticated psychological models and whose macro effects link to a system dynamics model. These hybrid models have great potential for addressing the needs identified in Chap- terÂ 2. Thus, we make this linkage between levels explicit. Although we do not have a separate chapter on integrated models, agent-based models and network models are discussed in Chapter 6, and the challenges of achieving such multilevel model integration are discussed in Chapter 8. Part II guide Chapter 3 discusses conceptual models and cultural models. Adequate conceptual models provide the foundation for development of computa- tional and mathematical models. Cultural models occupy a special position in our review because our interest is in understanding people at multiple levels of aggregation. The questions that concern us require the ability to model individuals, teams, communities, and entire societies. At each of these levels, cultural factors are at work, so we first explain what we mean by culture and cultural effects.
96 BEHAVIORAL MODELING AND SIMULATION The discussion next turns to a review of formal modeling approaches (Chapters 4, 5, and 6) organized using the four-part framework described above. For each modeling approach we describe the current state of the art, the most common applications of the approach, and its strengths and limitations for the problems described in Chapter 2, and we provide sug- gestions for further research and development. In Chapter 7, we turn to online games as a methodology. Gaming, the creation of an environment in which real people can play against one another or against artificial players, can be thought of as a methodology, but as these gaming models apply many of the other types of models, and as they involve people interacting with the games, we set them apart. Online gaming environments are both consumers of modelsâto create artificial players and the social effects of player actionsâand potential testbeds for generating data to develop and test models of the communications and actions of large numbers of individuals interacting in a simulated world. Chapter 8 discusses important methodological issues that are common across many modeling approaches, including modeling frameworks, tools, and data, and it includes a discussion of model verification and validation. Model validation is a key issue for complex social models, and we argue for a âvalidation for actionâ approach that considers how the model is to be used rather than attempting to evaluate model accuracy or model Âfidelity without considering context of use. Chapter 9 summarizes the state of the art in IOS modeling and its applicability to the requirements and uses dis- cussed in other chapters. References Borgatti, S.P. (2002). A statistical method for comparing aggregate data across a priori groups. Field Methods, 14(1), 88â107. DeJordy, R., Borgatti, S.P., Roussin, C., and Halgin, D.S. (2007). Visualizing proximity data. Field Methods, 19(3), 239â263. Johnson, J.C., and Griffith, D.C. (1998). Visual data: Collection, analysis, and representa- tion. In V. de Mjmck and E. Sabo (Eds.), Using methods in the field. Walnut Creek, CA: Altamira Press. Lave, C.A., and March, J.G. (1975). An introduction to models in the social sciences. New York: Harper and Row. Miller, J.H., and Page, S.E. (2007). Complex adaptive social systems: An introduction to com- putational models of social life. Princeton, NJ: Princeton University Press. Stokes, D.E. (1997). Pasteurâs quadrant: Basic science and technological innovation. Washing- ton, DC: Brookings Institution Press.