«Pandemonium in Silico: Individual Radicalization for Agent-Based Modeling Claudio Cioffi-Revilla and Joseph F. Harrison Center for Social Complexity ...»
Paper prepared for presentation at the Annual Convention of the International Studies
Association, Montreal, Quebec, Canada, March 16–19, 2011.
Pandemonium in Silico: Individual
Radicalization for Agent-Based Modeling
Claudio Cioffi-Revilla and Joseph F. Harrison
Center for Social Complexity and Department of Computational Social Science
Krasnow Institute for Advanced Study, George Mason University
Fairfax, VA 22030 USA
Abstract. How do individuals become radicalized, turning into terrorists, insurgents, violent actors. Computational agent-based models of irregular warfare, internal war, domestic political violence, and related conflicts require violent agents capable of carrying out attacks. Rather than introducing such agents as an exogenous process, as a Deus ex machina, this paper presents an agent-based model where radicalization is generated as an emergent phenomenon from within a population of individuals. The model (tentatively called “MASON RadicalAgent’’) is based on a new process-based theory of individual radicalization and is implemented in the MASON simulation system. Our paper describes the underlying theory, model structure, and some preliminary results intended for demonstration. This modeling effort is part of a broader project for modeling conflict in complex polities by combining computational simulations and network models.
1 Paper prepared for presentation at the Annual Convention of the International Studies Association, Montreal, Quebec, Canada, March 16–19, 2011.
Pandemonium in Silico: Individual Radicalization for Agent-Based Modeling Claudio Cioffi-Revilla and Joseph F. Harrison Center for Social Complexity and Department of Computational Social Science Krasnow Institute for Advanced Study, George Mason University How do individuals become radicalized, turning into terrorists, insurgents, or agents of related kinds of violence? Computational agent-based models of irregular warfare, internal war, domestic political violence, and related conflicts require violent agents capable of carrying out attacks against their targets. Rather than introducing such agents as an exogenous process, through some Deus ex machina, this paper presents an agent-based model (ABM) where radicalization is generated as an emergent phenomenon from within a broader population of individuals belonging to society. The model (tentatively called “MASON RadicalAgent’’) is based on a new process-based theory of individual radicalization that is implemented in the MASON simulation system. Our paper describes the underlying theory, model structure (specification and implementation), and some preliminary demonstration results for illustrative purposes. This modeling effort is part of a broader project for modeling conflict in complex polities by combining computational simulations and network models.
This paper contains five sections. In the first, we provide motivation and background. The second section describes how the MASON RadicalAgent model was developed in terms of specification and implementation. The third section presents a set of illustrative results. The fourth section presents a discussion of the model in terms of features, extensions, and future directions.
In real societies—past or contemporary—radicalized individuals are not born as radicals, nor do they just “come out of the blue.” Instead, radicalized individuals become that way through some 2 transformation process that takes them “from cradle to terror.”—and often “to grave” as well.
However, in most social or socio-natural ABMs where conflict or violence occur, the agents of violence are usually predetermined in the simulation (e.g., Epstein, 2002) or, alternatively, they may turn violent through some exogenous process. Some exceptions exist, as we discuss below, but most ABMs do not model radicalization endogenously.
The main challenge addressed in this paper is the endogenous production or emergence of radical violent agents as an emergent phenomenon generated within a broader population of agents— i.e., how to generate radical agents without hard-wiring them as individual extremists. There are basically two ways to address this challenge. The first is to draw on relevant social science (namely social psychology, behavioral science, and related social science of radicalization) to obtain the proper ideas for modeling radicalization. The second way is to rely on arbitrary algorithmic procedures that yield comparable results (pseudo-generative mechanisms, as these may be called), such as random assignation or other heuristic, non-theoretical procedures. Here we select the former strategy as a development in computational social science.
1.2 Background: Relevant extant literature
Insurgents, guerrilla fighters, rebels, terrorists, suicide bombers, violent anarchists, and similar kinds of extremist contentious actors have been implemented in social science simulation models since the early days of discrete dynamical systems (e.g., Ruloff’s (1975) DYNAMO model of guerrilla warfare in Afghanistan during the Soviet occupation). Other dynamical systems models based on similar aggregate frameworks include Akcam and Asal (2005), Madnick and Siegel (2007), and Wakeland and Medina (2010), among others.
In the area of agent-based models (ABMs), including those with an explicit territorial or geospatial component (spatial ABMs), contentious actors have been included in models since this area of computational social science began focusing attention on dynamics of political instability, regime fragmentation and change, civil war, secession, insurgency, and related phenomena. For example, guerrilla insurgency was implemented in the Iruba modeling project (Doran, 2005), which models a 32-province country afflicted by insurgency and related violent contentious politics between government forces and armed opposition. However, in the Iruba model radicalized agents exhibiting violent opposition are either created at initialization or they are “recruited”—i.e., they are not generated as emergent social process. Similarly, in Epstein’s (2002) civil violence models, individual agents behaving violently against government forces are not generated by any social or psychological mechanism. Rather, agents rebel based on grievance that is exogenously produced by random uniform values lacking systematic justification in social theory. The same is true of other models, such as Cioffi and Rouleau’s (2010) more recent RebeLand model, where agents turn radical when they become dissatisfied with their situation and as a result rebel against government; not because they undergo any psychological or cognitive transformation. This general theoretical deficiency in extant models of political violence does not detract from their contribution to a better understanding of the complex phenomena involved; but it does motivate development of models grounded in richer theoretical foundations.
3 In sum, to date there has not been an ABM where the phenomenon of individual radicalization is modeled as a psychological and cognitive process of transformation supported by viable social theory. Such a model would provide a number of advantages, such as a more systematic specification and implementation of radicalization—what generates insurgents, terrorists, guerrilla fighters, and other violent actors—in terms of known cognitive and psychological mechanisms.
2. Method: Model Development In this section we describe the procedure used to develop the RadicalAgent model in terms of specification, implementation, verification, and validation. We follow Sargent's (2004) and Cioffi's (2010a) criteria for verification and validation.
2.1 Model development: Specification and implementation The model was specified using Cioffi's (2010b) formal theory of individual radicalization and implemented in the MASON system (Luke et al., 2005).
2.1.1 Model specification The main specification objective for RadicalAgent was to create a model capable of occasionally generating radicalized agents based on a broader population of heterogeneous, bounded-rational agents (common people) undergoing situational changes in their lives. As summarized in Figure 1 and detailed in elsewhere (Cioffi, 2010b), individual radicalization is a process of cognitive and psychological transformation, not something individuals are born with or an occurrence “out of the blue.” The process of individual radicalization consists of three requirements (logically necessary conditions): Traumatic grievance, extremist indoctrination, and loss of killing inhibition. Roughly speaking, the first provides the motive for conducting violent attacks. The second provides cognitive support. The third disables murderous inhibitions humans have evolved as social members of a community.
In turn, each of these three events has a more detailed, fine-grained causal structure, down to so-called “leaf events” that mark the resolution of the tree. The overall causal mechanism is specified in terms of Boolean AND and OR connectives, corresponding to causal conjunctions and disjunctions, respectively. Formally, the occurrence of individual radicalization is specified in terms of a structure function expressed in terms of causal events linked by logic Boolean connectives. The emphasis here is on the use of the tree in Figure 1 for building the RadicalAgent model.
4 Figure 1. Backward-logic, conditional event tree for individual radicalization. The top, main event of interest is given by individual radicalization occurring as a “root event.” The root event will occur when the events below it have occurred, which depend on the events below them. The overall tree represents a theory of individual radicalization based on three requisite conditions linked by a first-level Boolean AND connective: traumatic grievance, extremist indoctrination, and loss of killing inhibition. Additional causal levels are specified by other AND and OR connectives until reaching “leaf events” that mark the resolution of the tree. Source: Adapted from Cioffi (2010b).
Model specification consisted of several steps. First, we identified the leaf events in the success
tree for individual radicalization. These events include, for instance:
Second, we identified a set of events that actually occur as part of a simulation process. (The simulation model in which RadicalAgent is inserted may be an existing model or some other
ABM.) Examples of this other set of events include:
● Economic transactions. Sales of goods, transfers of money, returns on investments.
Some of these can cause economic loss.
● Political events. Campaign events, elections, appointments. Some of these can cause social loss.
● Loss of economic status. An individual undergoes significant losses in the course of work or other economic transactions.
● Loss of social status 5 ● Migration of agents. Movements of population due to push or pull factors, such as conflict, natural disasters, attractive opportunities.
● Other events.
The individual grievance model can be viewed as a System Dynamics model. The level of grievance is occasionally increased by traumatic events and decreases via a decay process that can have various forms: Exponential, linear (constant decay), parabolic (time-since-last-trauma decay), and sigmoidal, each with a characteristic half-life (t such that the level of grievance is half the initial value G0). Grievance events can be generated stochastically or from a file containing a stream of events. Figure 2 shows examples of these four patterns of decay driven by regularly spaced stochastic shocks of varying intensity. Stochastic generation of grievance events enables simulation of alternative scenarios using various assumptions, although an exponential distribution (Poisson) of time-between-events is common in many social systems and processes (Bartholomew, 2005; Cioffi, 1998: 52, table 2.1).
Figure 2. Level of traumatic grievance (y-axis) as a function of time (x-axis), given four different decay functions (linear, exponential, parabolic, and sigmoid) under an identical set of time-distributed shocks.
Adapted from Harrison and Cioffi (2010).
An example of the echo mechanism is shown for four settings in Figure 3, assuming four different (illustrative) social structures. The simplest result is the 4-set, 1 echo case, where an individual agent’s traumatic grievance is felt by four local von Neumann neighbors (Figure 3a).
In the most complex case, 8-set, 2 echoes, the effects are felt throughout a broader community (Figure 3d).
(a) 4-set, 1 echo (b) 4-set, 2 echoes (c) 8-set, 1 echo (d) 8-set, 2 echoes Figure 3. Echo effect of traumatic grievance. In (a) the center agent suffers a traumatic grievance which is echoed to its von Neumann neighbors at 50% off its original intensity. In (b) the event is echoed to the von Neumann neighbors, then echoed again to each of their neighbors, including back to original agent. (c) and (d) show the same process but using the Moore neighborhood instead. Adapted from Harrison and Cioffi (2010).
The second requirement in the process of individual radicalization (Figure 1) is indoctrination into an extremist belief system. Indoctrination was modeled as formation of extremist views in an opinion dynamics model, building on Jager and Amblard (2005). As part of this we replicated the Jager-Amblard model in MASON (Harrison and Cioffi, 2010), successfully reproducing patterns of opinion polarization under a variety of conditions.
Extending the Jager-Amblard model, we also introduced demagogues (extremist preachers) in the social network. Demagogues are akin to signal transmitters of extremist views with unchanging opinions and capacity to instigate further extremism.
The procedure just described resulted in an agent-based model where individual radicalization is generated by endogenous iterations, not hard-wired into the model. Emphasis here is on the more complex, multi-level causality of the grievance component, as opposed to the indoctrination and killing inhibition components of the model in Figure 1.
72.2 Model verification