MODELING GENOCIDE 1 Modeling Genocide at the System and Agent Levels Elizabeth M. von Briesen, N. Gizem Bacaksizlar, Mirsad Hadzikadic University of North Carolina at Charlotte
MODELING GENOCIDE 2 Abstract This research works to bridge the gap between the knowledge of social scientists in the field of genocide studies, and that of computational experts who have the ability to produce powerful models and simulations that can assist in the study and prediction of genocide. A thorough understanding and visualization of the system level dynamics of the problem inform the implementation of an agent-based model that simulates local interactions in a society composed of two identity groups. The model explores how dynamics between agents with different identities, ideology, influence, susceptibility, and threshold to act against those of a different group, lead to the emergence of genocide. Early results indicate the model s usefulness in exploring these dynamics and revealing the underlying factors of importance. The majority of simulations quickly result in genocide. However, the occasional appearance of sustained environmental stability points to the model s strong potential for revealing significant differentiating factors between a stable society and one that experiences genocide. Given further refinement, this study has the potential to become a generic model for social scientists and genocide researchers that is useable, interpretable, and useful for predictive and analytic purposes. Keywords: genocide, system dynamics, agent-based modeling, complex systems
MODELING GENOCIDE 3 Modeling Genocide at the System and Agent Levels The phenomenon of genocide is complex, as is the task of deriving its definition. Verdeja (2010) states that [u]ntil scholars reach some degree of consensus on the definition of genocide, it will be difficult to compare theories in any rigorous manner. While the above job falls on the shoulders of social scientists and the entities with which they work, it is important in the context of this computational effort to have clear direction. Verdeja does an excellent job of surveying a number of related works, and points out what he and others see as flaws in the UN definition of genocide a primary issue being its reference only to ethnic, racial, national, and religious groups. He points to the lack of recognition of political, economic, and other groups as one of many weaknesses, with another being the lack of a threshold which would indicate the level at which mass killing constitutes genocide. Computational modeling generally requires a simple foundation, and given the lack of consensus, Straus (2012b) definition of genocide as a large-scale, organized, groupdestructive violence that targets a specific social group in a territory is sufficient. This is a general statement, but its simplicity is appropriate in the domain of computational research. Creating a robust and flexible model that produces interpretable results requires stripping away details and focusing only on the most basic building blocks necessary. Given the above definition, the next task is to derive a satisfactory framework from which to build a simulation, and a system (or macro) level perspective is a logical starting point. This research attempts to gain new insight and understanding into the fundamental drivers of genocide in order to better inform preventative policy decisions. Complexity theory guides the analysis, which begins by exploring the problem at the macro level. This approach produces a broad picture of the large-scale dynamics of genocide. However, given complexity
MODELING GENOCIDE 4 theory s focus on local interactions, the model implementation uses a simplified framework in an agent-based model for initial exploration of the problem. Method System Level View A survey of the work of major scholars in the fields of genocide studies shows that the complexity of the problem framework increases with the depth of the research. The causal loop diagram from the field of System Dynamics provides a useful structure for visualizing the problem framework at the system level. Figure 1 is a simplified causal loop diagram created with Vensim PLE Software (Version 6.3G; Ventana Systems, Inc., 2006). It shows a number of relationships and feedback mechanisms. Each variable has a descriptive name, and is connected to others by arrows of specified direction and polarization. A positive sign () indicates a positive relationship. This means that an increase in the variable at the beginning of the arrow leads to a corresponding increase in the variable at the end of the arrow, and vice-versa. A negative sign (-) denotes the opposite, with an increase at the beginning of the arrow corresponding to a decrease at the end of the arrow, and vice-versa. A positive/negative sign (/-) indicates uncertainty in the nature of the variable relationship. Feedback loops within the diagram show the dynamics of change over time between sets of connected variables.
MODELING GENOCIDE 5 Allied Assistance to the Elite Regime External Military Intervention Against the Elite Regime War Sanctions Subnational Actors R3 Threat to the Elite Extreme Elite Ideology R1 R2 /- Genocide Against the Out-Group Out-Group Armed Resistance /- /- In-Group Civilian Compliance- /- - - Societal Diversity - Strong Economy Political Resistance Preexisting Democratic or Inclusive Ideals Figure 1. Simplified diagram for effect of Extreme Ideology on Genocide Against the Out-Group. Extreme Elite Ideology and In-Group Civilian Compliance. The central loop (R1) connects two sets of primary actors. The elite represent the ruling class of the dominant societal group, and the in-group represents civilian actors who are of the same identity as the elite. The dynamic shown between the two is specific to the incidence of genocide. While there is not complete consensus among scholars that the presence of an extreme ideology on the part of
MODELING GENOCIDE 6 the elite is a required condition for genocide to occur, it is often acknowledged by domain scholars to have critical importance. Waller (2001) states that some scholars contend that the lethal soil of specific extraordinary cultures, nurtured by the influence of an extraordinary ideology, is the origin of extraordinary human evil. Verdeja (2002) identifies [a]rticulation and repetition of exclusivist ideology by elites and espousal by the public as a substantial contributing factor to genocide. Straus (2012a) also highlights ideological paradigms as a driver of genocide, where the central mechanism is exclusion. As shown in the diagram, this dynamic relationship is characterized by In-Group Civilian Compliance in the presence of an Extreme Elite Ideology through a positive effect in both directions. This creates a reinforcing feedback loop (R1). While not yet fully validated in this research, it is logical to conclude that in order for an elite regime to commit genocide, they require either the active participation of their civilian population, or its tacit approval. The internal dynamics of the civilian group are not clearly distinct in one direction. Waller (2001) and Straus (2012b) both point to the possibility of support or restraint originating from the ingroup. Waller states that [t]here is a diffusion of responsibility within groups that can make evildoing a relatively simple matter. In addition, groups have power to repress dissent and, thus, encourage the abandonment of the individual self. However, he contrasts this viewpoint by explaining that groups also have positive potential by saying that groups can even provide the security to oppose potentially destructive ideas and practices. Straus states that while ideology is often linked to escalation of mass violence, it can also be a source of restraint. The above variation is captured in the diagram through the rise and fall of In-Group Civilian Compliance. Genocide Against the Out-Group and Societal Diversity. Further extending the diagram is Genocide Against the Out-Group, where the out-group is of a different identity than
MODELING GENOCIDE 7 the elite and its civilian population. An increase in extremist elite ideology leads to an increased likelihood of genocide, and by extension decreases Societal Diversity. Cox (2017) states that genocide is enduringly destructive: The annihilation, partly or wholly, of a group diminishes the diversity and richness of the human race. When a people is eradicated or dispersed and its traditions and culture erased, all of human civilization loses much that can never be regained (p. 189). While this research does not yet include a full validation of the connection of a decrease in diversity to an increase in civilian compliance, it is probable that its effect should be included such that it closes another reinforcing loop (R2). Subnational Actors. Connected to the elite are what Straus (2012a) refers to as subnational actors. He defines them as a mix of important actors province- and town-level civilian administrators or security forces; influential professional, religious, or business actors who shape policy in rural areas; or ethnic groups that are located on the periphery. He then asks: Are alliances between national and local actors necessary for genocide to occur? Are policies of genocidal mass violence accelerated or initiated at the local level? In short, in what ways do subnational dynamics shape genocide? Due to the theoretical importance of subnational interaction with elite actors, they are included in the diagram through a reinforcing feedback loop (R3). Further research results from Straus and others will help better validate and define what is likely to be an important feedback mechanism. War. The next extension of the diagram is through the inclusion of factors of restraint and escalation (Straus, 2012b). The majority of scholars in this domain agree that genocide predominantly occurs in the presence of war or other armed conflict. Verdeja (2002) states that war, coup, revolution, economic collapse or any other equally rapid and catastrophic social transformation are common structural elements preceding genocide. These conditions greatly
MODELING GENOCIDE 8 increase instability in a system, which the elite can then exploit to drive divisive and genocidal agendas. Straus (2012b) states that the use of violence is legitimated in war and that coupling this violent instability with exclusionary ideology explains two important factors of escalation. Threat to Elite and Genocide Against the Out-Group. In the diagram, War has a positive effect on Threat to the Elite, which is in turn has a positive effect on Extreme Elite Ideology. While this research has not yet fully validated the above, the existence of a threat to the elite regime is likely to be an important factor of escalation. Straus (2012a) provides a concise description of this dynamic by stating that war creates threat and insecurity, which in turn increase the probability that violence will be used to counter the threat and that this is the core of a strategic perspective that most authors share. The ambiguity of the diagram s labeled effect of War on Genocide Against the Out-Group highlights the possibility that armed conflict may have a direct positive or negative effect on genocide. Territorial advances by external actors could reduce the incidence of genocide by providing safe havens for out-group civilians, although the opposite is possible in the presence of territorial advances made by the extreme elite regime. The rise and fall of Nazi Germany in WWII, and the corresponding effect on its genocidal activities provides an obvious example. Further research will help better define, characterize, and validate these dynamics. War Other Factors. There are three additional variables in the diagram that have effects on War. It is logical to conclude that Allied Assistance to the Elite Regime has a positive effect, however, in the cases of External Military Intervention Against the Elite Regime and Sanctions, the relationships are more difficult to define. Either could be a factor of escalation or
MODELING GENOCIDE 9 restraint depending upon conditions unique to the circumstances under examination, and this is an area that requires further research. Internal Societal Factors. The scope of the diagram extends to include several variables internal to the society in question. All are connected directly to In-Group Civilian Compliance, as they have specific mechanisms that directly affect this set of actors. Each is characterized below, and future research here will more thoroughly explore all variables and their relationships. Out-Group Armed Resistance. In addition to its ambiguous effect on In-Group Civilian Compliance, Out-Group Armed Resistance has a direct and positive effect on Threat to the Elite as shown in the diagram. This relationship in turn affects the level of extremism in elite ideology and other connected variables according to their polarity. As indicated above, is less clear how Out-Group Armed Resistance affects In-Group Civilian Compliance. The early theory of this research is that if the resistance is effective, civilians may be less likely to passively accept the extremist ideology of the elite, and vice-versa. This relationship may be difficult to validate, but is certainly worthy of discussion. Strong Economy and Political Resistance. Straus (2012b) discusses the dynamics of mass violence and economics, and includes in his analysis a focus on the middle class. He states that [m]iddle-class and economic elite actors have an incentive to seek stability, because largescale disruption and violence can threaten their wealth and property. They also are presumably influential; their opinions will matter to the ruling elite. In the diagram, Straus theory provides initial justification for a Strong Economy to positively affect Political Resistance, which in turn may have a negative effect on In-Group Civilian Compliance. Again, more research is required
MODELING GENOCIDE 10 here, but Straus work is compelling in highlighting the importance of considering these dynamics. Preexisting Democratic or Inclusive Ideals. Finally, the diagram includes Preexisting Democratic or Inclusive Ideals as a variable that has a negative effect on In-Group Civilian Compliance. Straus (2012b) explores both positive and negative cases of genocide in sub- Saharan Africa from 1960 2008. A negative case is specified by circumstances that made genocide appear to be a likely outcome, but the situation resolved without its occurrence. Among other factors, Straus finds that [n]o genocide took place in any full-fledged democracy. The initial explanatory theory of this research is that a preexisting set of democratic ideals restrains civilian compliance. If these types of ideals are firmly entrenched in a society, they may cause its people to be less compliant in the face of extreme ideologies both from within their ranks, and from the elite. This completes the initial overview of the scope of the problem of genocide as viewed from the system level. However, any study of a human social system will, by default, include nonlinearity. This research narrows its focus on the problem through the lens of complexity theory, using the causal loop diagram of the system to determine an optimal starting point. Agent Level View A complex system is one composed of interacting parts in which those interactions sometimes lead to non-linear, emergent, system level properties. When examining such systems, the researcher must select an appropriate level of granularity. In the case of genocide, one could model only the internal psychological nuances of an individual human being in the face of extremism, and at the other end of the granularity spectrum one could model the dynamics of states as they interact with other states. Regardless of the frame of reference, local interactions
MODELING GENOCIDE 11 are the primary focus in the study of a complex system. Emergent properties are a critical observation, but understanding their causes requires a thorough grasp of micro-level interactions and dynamics. Genocide emerges from a system of people interacting with other people. The system level view as shown in Figure 1 informs the next step, which is an implementation of an Agent- Based Model (ABM). In this case, the system could represent anything from a nation-state to a local population of a village, but the more important point of focus is on the nature of human interactions. Figure 2 shows the inner feedback loops of the diagram in Figure 1. This section of the diagram helps visualize a very basic set of dynamic interactions and effects. The ABM implemented in this research simulates the dynamics between elite and civilian actors with respect to ideology and compliance in order to better understand genocide s lowest level origins. Extreme Elite Ideology R1 R2 Genocide Against the Out-Group In-Group Civilian Compliance- Societal Diversity - Figure 2. Focal points for an agent-based model.
MODELING GENOCIDE 12 Model Implementation. This model is implemented using NetLogo (Version 5.3.1; Wilensky, 1999). Structuring an ABM requires clear definitions of the environment, agent attributes, and rules for interaction. Each is defined below. Environment. As stated above, the environment of the model is a closed system of people in that there are no external forces in the simulation to which agents respond. All agents have random initial positions and subsequent movement within the environment. Agent Attributes. Agents are endowed with six attributes: Identity (Group A or B). In order to simulate identity-based conflict, agents are divided into two groups and have a corresponding Identity attribute. A division into two groups represents the lowest level of societal division, and are thus the optimal starting point. Ideology (Dynamic, 0 1, random initialization): A higher Ideology level indicates that an agent has a more extreme view of agents from the opposing group. A review of genocide studies literature indicates that an ideological division of a society within the minds of its members is a necessary condition preceding genocide. Waller (2001) states that, we tend to see in-group members as very similar to us and out-group members as all alike and very different from us. Without much prompting, we then extend these assumptions to a belief that our group is better than the out-group. The model does not specify an out-group or an in-group, and if all other attributes and environmental conditions are held equal, an agent s Ideology setting simply simulates its perception of the other. No group has an advantage or position of dominance. Influence (Static, 0 1, random initialization): In the model, the elite are simulated through the Influence attribute. According to the rule structure outlined below, an influential agent has a higher likelihood of causing its neighboring agents to increase their Ideology level.
MODELING GENOCIDE 13 This is a highly simplified representation of the elite in a society, and the initial model gives all agents the same random chance of being an influential, and therefore elite, actor. Susceptibility (Static, 0 1, random initialization): This attribute determines to what degree an agent updates its Ideology in the presence of a more influential agent of the same Identity. Its random initialization allows an equal distribution throughout the agent population, simulating the wide variety of personality differences possible in human societies. Threshold to Act (Static, 0 1, random initialization): The threshold attribute sets the level beyond which an agent acts against other agents in its vicinity with a different Identity. In the historical record, not all genocides have involved a high level of civilian participation. However, as discussed above, their compliance is necessary in order for the elite to further a genocidal agenda. In the model, acting against another agent can simulate active participation or mere compliance in the face of genocidal acts of others from the same identity group. Radius of Sight (Static, 1 50 patches, global setting): In the environment, agents are located on patches, and agents of both groups have the same level of local visibility from their position as determined by this attribute setting. In order to avoid bias, the model has the same rule structure for all agents. While it is possible to add a wide variety of interaction rules, the model includes only the simplest possible set of rules, leaving layers of complexity to future implementations. Interaction Rules. These rules form the foundation for agent adaptation. Their simplified nature makes it possible to more easily identify micro-level sources of emergent properties. 1. Agent n identifies the most Influential agent with its same Identity in its radius of sight.
MODELING GENOCIDE 14 2. If there exists such an agent, and its Influence is higher than that of Agent n, it is identified as Agent x. Agent n updates its Ideology according to the following formula: (New Ideology) Agent n = (Old Ideology) Agent n [Ideology Agent x (Old Ideology) Agent n ] * Susceptibility Agent n. This means an agent s change in Ideology when in the presence of a more influential agent from its own group is directly proportional to its Susceptibility. 3. Next, Agent n checks to see if (New Ideology) Agent n > (Threshold to Act) Agent n, and if this is true, Agent n kills all agents from the other identity group in its radius of sight. 4. Finally, every agent has an equal probability of reproducing in each model cycle. A new agent only inherits the Identity attribute of its parent. All other attributes are initialized randomly according to the same structure specified above. These attributes and rules are the minimum set required to simulate genocide in an ABM in the context of the system view outlined in Figure 2. While it was possible to choose a different starting point, the works of Waller (2001) and Straus (2012a, 2012b) greatly informed the initial model components. Straus helped define the system level understanding, and Waller provided support for key elements of Straus model, such as the presence of an extreme ideology. Waller also highlighted the importance of the psychology of the individual, and how it drives ordinary people to commit extraordinary evil. His focus on the psychological dynamics in the environmental presence of an extreme ideology with respect to the other guided the selection of an ABM with the above parameters for this initial implementation.
MODELING GENOCIDE 15 Early Results A broad exploration of model settings leads to two interesting and informative outcomes. For all model runs, initial populations are set to 50 agents from each identity group, all agents have a 5 patch radius of sight, and all have a 1/10 probability of reproducing in every cycle. Outcome 1 In all runs using the above conditions, one agent group eventually dominates the other such that only its agents survive. Figure 3 shows the most common scenario. (Note that this NetLogo environment is a torus, meaning the edges wrap around to the opposite side.) Here blue agents greatly outnumber green agents in only 137 cycles (or ticks), and have also surrounded those agents on all sides. When this happens, the surrounding agents quickly become the only surviving identity group. In this model run, all green agents are killed in 284 ticks. Figure 3. Agent group isolation.
MODELING GENOCIDE 16 With all parameters equal for both populations, including random initial placement, movement, and reproduction, genocide always emerges at the system level. The only requirement is that agents to update their ideology in the presence of a more influential neighbor (of the same identity) according to the original agent s level of susceptibility. Note that Ideology can increase or decrease for the agent in question, and that the mere presence of the concept of the other as a group against which the agent with sufficiently extreme ideology will act is all that is required to quickly isolate and exterminate one identity group from the environment. Outcome 2 While outcome shown below in Figure 4 occurred less frequently than the scenario from Figure 3, it consistently led to model runs lasting more than 1000 ticks. Its defining feature is that each agent population establishes a continuous border in the environment. This appears to provide temporary stability for both identity groups, as neither can easily break through the continuum of agents of differing identity along that border. The model run shown in Figure 4 lasted 1165 ticks before only the green population remained. In this scenario, it is typically the breach of the continuum, by one group or the other, that brings about genocide.
MODELING GENOCIDE 17 Figure 4. Continuous population distribution. The continuous population distribution outcome can also be interpreted as simulating ethnic cleansing. Both the blue and green populations have established environments that are void of the other identity type, and as long as they can maintain a continuous border, they cannot be easily exterminated. Conclusions and Future Work These preliminary results are encouraging. A full exploration of the model parameters will likely reveal critical information about the local level dynamics that lead to genocide. While the model is simple, it is powerful enough to demonstrate possibilities for policy recommendations. Further exploration of the scenario leading to agent group isolation (Figure 3) should begin with the inclusion of additional identity groups. The genocidal outcome here is expected, which leads to the question of how further fractionalization of the society along identity lines
MODELING GENOCIDE 18 impacts the result. Does the addition of a third identity bring forth stability in the environment? How does the initial size of that population affect its impact on the environment as a whole? How do the dynamics change when adding additional groups of varying sizes? Each of these questions addresses how fractionalization affects stability in the presence of identity-based division and provides a basis for future research. The continuous population outcome (Figure 4) has a strong potential to inform policy related to genocide. The emergence of genocide in this scenario is quite different from that seen in the agent group isolation outcome, as evidenced by a much longer period of stability before breakdown. Identifying the keys to the difference could provide valuable information about the core mechanisms of genocide. Is something significant happening on the boundaries dividing the two populations? Do agents along the boundaries have characteristics that are different, on average, than those in the central region of their identity group? What are the differences and similarities in the characteristics of boundary agents between outcome scenarios? How do agent and system conditions change immediately preceding the breakdown of stability and ensuing genocide? All of the above are intriguing questions, and require a very thorough analysis of the model in order to obtain definitive answers. Those answers have the potential to impact the policy discussion around genocide, as they provide insight into the local dynamics and levers that support stability in a fractionalized society. Finally, this implementation of the model uses a random distribution for all parameters. While this has been informative, there may be others, such as normal or power-law distributions, which could produce insightful results. A full analysis of the model requires testing different distributions on every parameter and combination of parameters in order to determine differences in outcomes.
MODELING GENOCIDE 19 Social science researchers in genocide studies have a significant restraint on their ability to analyze, understand, and predict the problem because they must wait for an emergence, one that brings with it a great deal of human suffering, in order to validate their models. This research attempts to synthesize their knowledge, adapting it to a computational approach in order to develop a model that is useful and informative. While the research and model are not fully developed, they have already produced intriguing results. As this work matures, it has a significant potential to put the power of computational modeling and simulation into the hands of social scientists such that they can benefit from its analytic and predictive capabilities. This in turn increases the ability of society as a whole to more effectively address and mitigate the problems associated with identity based conflict.
MODELING GENOCIDE 20 References Cox, J. M. (2017). To kill a people : Genocide in the twentieth century. New York, New York: Oxford University Press. Straus, S. (2012a). Destroy them to save us: Theories of genocide and the logics of political violence. Terrorism and Political Violence, 24(4), 544 560. http://doi.org/10.1080/09546553.2012.700611 Straus, S. (2012b). Retreating from the brink: Theorizing mass violence and the dynamics of restraint. Perspectives on Politics, 10(02), 343 362. http://doi.org/10.1017/s1537592712000709 Ventana Systems, Inc. (2006). Vensim PLE Software (Version 6.3G) [Software]. Available from http://www.temoa.info/node/3682 Verdeja, E. (2002). On genocide: Five contributing factors. Contemporary Politics, 8(1), 37 54. http://doi.org/10.1080/13569770220130112 Verdeja, E. (2010). Genocide: Clarifying concepts and causes of cruelty. The Review of Politics, 72(03), 513 526. http://doi.org/10.1017/s0034670510000343 Waller, J. (2001). Perpetrators of genocide : An explanatory model of extraordinary human evil. Journal of Hate Studies, 52, 5 22. Wilensky, U. (1999). NetLogo (Version 5.3.1) [Software]. Available from http://ccl.northwestern.edu/netlogo/