IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 44, NO. 1, JANUARY 2014 103

Social Network Modeling and Agent-BasedSimulation in Support of Crisis De-Escalation

Michael J. Lanham, Member, IEEE, Geoffrey P. Morgan, and Kathleen M. Carley, Senior Member, IEEE

Abstract—Decision makers need capabilities to quickly modeland effectively assess consequences of actions and reactions in cri-sis de-escalation environments. The creation and what-if exercisingof such models has traditionally had onerous resource require-ments. This research demonstrates fast and viable ways to buildsuch models in operational environments. Through social networkextraction from texts, network analytics to identify key actors, andthen simulation to assess alternative interventions, advisors cansupport practicing and execution of crisis de-escalation activities.We describe how we used this approach as part of a scenario-drivenmodeling effort. We demonstrate the strength of moving from datato models and the advantages of data-driven simulation, which al-low for iterative refinement. We conclude with a discussion of thelimitations of this approach and anticipated future work.

Index Terms—Computer simulation, information diffusion,social network analysis, text mining.

I. INTRODUCTION

E FFECTIVE crisis response requires thinking through theimplications and interactions of complex sets of events,

an error-prone process for humans with high-stakes in the de-terrence domain. War games, and modeling and simulations(M&S) in support of war games, can mitigate lack of experi-ence and support forward thinking by providing safe venues toassess alternatives. Simulation effort intended to support for-ward thinking, however, often has long time cycles to develop;the need for the tool is overtaken by events before the tool isready. This paper presents a rapid metanetwork (multimode,multilink) modeling approach using concept extraction tech-niques to develop models to examine scenarios within a usefultime span.

We present and discuss our process to rapidly develop usefulmetanetwork and information diffusion models through semi-automated analyses of text corpora; how we applied the ap-proach to deterrence and crisis de-escalation scenarios, and ourlessons learned. We discuss how our method’s outcomes were

Manuscript received May 20, 2012; revised October 1, 2012; acceptedNovember 12, 2012. Date of publication August 8, 2013; date of current versionDecember 20, 2013. This work was supported in part by the Office of NavalResearch (ONR) Contract N00014-08-11223 and by the Air Force Office ofSponsored Research under Grant 600322. The views and conclusions containedin this document are those of the authors and should not be interpreted as repre-senting the official policies, either expressed or implied, of the ONR or the U.S.government. This paper was recommended by Associate Editor K. M. Sim.

The authors are with the Center for Computational Analysis of Social andOrganizational Systems, Institute of Software Research, School of ComputerScience, Carnegie Mellon University, Pittsburgh, PA 15213 USA (e-mail: mlan-ham@cs.cmu.edu; gmorgan@cs.cmu.edu; kathleen.carley@cs.cmu.edu).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TSMCC.2012.2230255

triangulated using a multimodeling approach and offer caveatsand potential future work.

II. RELATED WORKS

Conflict reduction is an often-researched area of humanknowledge. Indeed, there are entire journals dedicated to thestudy of international conflict: The Journal of Conflict Resolu-tion; The Journal of Conflict & Security Law; Peace, Conflict,and Development; The International Journal of Conflict Man-agement among others. Assessment of conflict reduction andde-escalation effort can be as simple as “is there no longer ashooting war” to much more nuanced sets of measures of effec-tiveness and measures of performance. Use of computer-aidedM&S has ranged from human-based experiments [1] to effortsto include environmental and cultural framing to contextualizeinformation [2] as well as correlation models [3]. Richardsonintroduced a purely mathematical set of models in [4], whileRuloff used system dynamics to model international relationsscenarios in [5]. Yilmaz et al. present a summary of modelingeffort of the past 30 years from game theory to early socialphenomena models in [6].

Building models from unstructured data also has a wide-ranging application history, from cell tower data for social net-work inference [7] to developing emergent ontologies [8]. In-deed, a grand vision of the “Semantic Web” was to bring struc-ture and computable meaning to the World Wide Web [9], [10].Social network construction from web-based sources has beendone from the mid-90s to the current day through the use of websearch tools’ application programming interfaces [11], [12]. Wediffer from these methods because the constructed networks areboth analyzed and serve as inputs into M&S environments. Likethe research works in [13] and [14], we need to rapidly buildM&S capable models, but we use sociolinguistics theory andmachine-learning-based topic modeling, as introduced in [15],to provide a model construction mechanism.

Belief modeling plays a key role in the effort, although weare not using a belief–desire–intention modeling paradigm [16].Modelers have used beliefs to support individual goal-orientedbehaviors [17] as well as simulating threshold-based behaviors[18]. Beliefs, by which we mean attitudes, convictions, andopinions, do not require exposure to or awareness of knowledgeand facts for people to sustain them.

Friedkin developed a belief modeling method with his so-cial influence theory in [19] that brought four innovative con-cepts to social theory and modeling: 1) he relaxed previous as-sumptions where agents had to either conform or deviate froma fixed consensus (the public choice model); 2) his methoddid not have to lead to consensus, and could support stable

1094-6977 © 2013 IEEE

104 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 44, NO. 1, JANUARY 2014

patterns of disagreement; 3) he provided a multilevel the-ory where microlevel cognitive processes could influence andconstrain macrolevel changes; and 4) his methods supportedquantitative analysis of the systematic consequences of socialstructures.

The agent-based modeling performed in this effort makes di-rect use of the Construct information diffusion simulation devel-oped at Carnegie Mellon University’s Center for ComputationalAnalysis of Social and Organizational Systems. Construct hasits roots in constructuralism and combines structuralism withsocial influence theory [20], [21]. Constructuralism, in brief,asserts that agents’ actions, perceptions of selves and others,learning and forgetting knowledge and beliefs are all constantlyinfluenced by the agents’ surrounding environment, particularlytheir surrounding social environment.

We use Automap to demonstrate a fast method to build simu-lation models for use in Construct, what we refer to throughoutthis paper as the Data to Model (D2M) process. These modelsenable practicing crisis de-escalation and deterrence. They alsoallow policy analysts to evaluate multiple counterfactual sce-narios. We use Construct as a belief diffusion model, arguingthat if policy makers believe they “should” go to war, then thedeterrence calculus has failed. Construct is a validated model ofbelief and knowledge diffusion [22]–[24] that is shown to fit awider range of data than reinforcement theory and informationprocessing theory [25].

The multimodeling component of this paper is limited to adiscussion of the simultaneous development and use of threedifferent modeling tools, each with very different origins andtheories of function to arrive at congruent results.

III. DATA-TO-MODEL PROCESS

The data-to-model process (D2M) we use is a systematic,computer-assisted, repeatable approach with these steps [26],[27]:

1) collect data;2) clean the text corpus;3) ontological cross classification; and4) generate static data for analysis.Collecting data is the first step. The D2M process focuses

on the challenges associated with unstructured data, althoughother forms of data can contribute to later analysis. We convertlarge amounts of unstructured texts into rich multimode, mul-tiplex, and multilevel relational networks (i.e., metanetworks)for use in dynamic simulations. The second step in our processis cleaning the text corpus. Text data, like all language, are rifewith ambiguity. Data cleaning removes and/or clarifies redun-dant or ambiguous references, removes noise words, performspronoun resolution, and acronym disambiguation. Step three isontological cross classification; this step classifies phrases, forexample, “President” is classified as an agent, as well as resolvesambiguities when words have two meanings, such as “battery,”which can be either a resource or an agent. Illustrative classesare agents, knowledge, and tasks [28], [29]. Typical semanticsof the networks between them are shown in Table I. An ana-lyst or planner iterates through steps two–four as many timesas appropriate to the demands of their leadership. Ideally, she

TABLE ISIMPLE METANETWORK COMPRISED OF SIX NETWORKS FROM THREE

TYPES OF NODES

Fig. 1. Sample Multimode network of agents (circles, multicolored by coun-try) and knowledge (hexagons, red), sized by Eigenvector Centrality.

would maintain up-to-date models through periodic additionsto her corpora with new information and sources. The D2Mprocess creates intermediate artifacts, allowing the process tobe run without modification on new data or tweaked to improvethe resulting model(s). Improvements can be subjective in theeyes of subject matter experts (SMEs), objective with respect toleader-specified network analytics and metrics or a combinationof the two.

The final step in the process, generate static data for analy-sis, identifies linkages among the nodes through windowing, i.e.,through proximity of the cleaned nodes in the text. These link-ages are across multiple modes, creating a metanetwork such asthat seen in Fig. 1. The analyst can then use this metanetworkfor point-in-time analysis as well as input to simulations—inour case, a diffusion simulation. An important difference be-tween these networks, and traditional network science’s focuson agent-by-agent interactions, is the inclusion of the nonagentnode classes in the networks and in the analysis [30].

These four steps are not, themselves, unique. What is unique,from searches of related literature, is the use of these four stepsin an operational-environment-friendly, low barrier-to-entry andlow maintenance cost process of text mining to build multi-mode social networks that become the inputs into agent-basedsimulations. Fig. 2 depicts this larger process in an abbreviatedflowchart.

IV. BELIEF FORMATION IN CONSTRUCT

Construct is a widely validated, agent-based model, with afocus on information diffusion and belief change [22]–[24],[31], [32]. Agents interact with those with whom they are sim-ilar (e.g., homophily) [33], which is a proven cross-cultural

LANHAM et al.: SOCIAL NETWORK MODELING AND AGENT-BASED SIMULATION IN SUPPORT OF CRISIS DE-ESCALATION 105

Fig. 2. Flowchart of this paper’s process.

phenomenon [34]. Agents also interact with those from whomthey seek information they do not have (e.g., expertise seek-ing) [32], [35]. Agents exchange and learn correct and incorrectinformation (implemented as vectors of 0/1 bits that we refer toas knowledge bits) as well as exchange information about ego’sand alters’ beliefs [21], [25].

In Construct, agents’ beliefs may be anchored toknowledge—sets of knowledge bits can contribute positive ornegative valence for a belief and each agents’ belief values rangefrom [−1.0, 1.0]. For this effort, we used knowledge-anchoredbeliefs only. Belief formation, on a per-agent per-turn basis, is asummative function between an agent’s prior beliefs mitigatedby their ability to be influenced by their alters (extended fromFriedkin in [19]) and their similarity to their alters magnified bytheir ability to be influenced by their alters. We will build to thisformal equation (12) in the following paragraphs as Constructimplemented it in [36] and [37].

Agents in Construct also have error-prone perception of who-knows-what and who-believes-what. We use the term transactivememory for this perception. Construct implements transactivememory as a 3-D binary matrix, which is denoted as knowledgetransactive memory (KTM), with indices i, j, and k where agenti (the ego) perceives that agent j (the alter) is in possession ofknowledge bit k. The same convention applies to the belieftransactive memory (BTM) for each belief b in the simulation.The expressions are shown as follows:

∀i, j ∈ Agents (A)

∀k ∈ Knowledge Bits (K)

∀b ∈ Beliefs (B)

KTMijk and BTMijb (1)

As alluded to earlier, homophily preference is driven by ameasure of knowledge similarity (SK) and belief similarity (SB).As shown in (2) and (3), using the expressions in (1), it is thesum of self-perception per-bit multiplied by the perception ofeach connected agent’s per-bit knowledge or belief

SKij =∑

k(KTMiik × KTMijk ) (2)

SBij =∑

b(BTMiib × BTMijb). (3)

The ability of an agent (ego) to affect its connected neigh-bors (alters) is called social influence (RS). RS is a function ofconnectedness between an ego i and its n alters as well as SKand SB. It is shown in (4) and incorporating (2) and (3). α in(4) is an exogenous parameter that is the weight an agent places

on SK versus SB. In this experiment, we set α to 0.50 for equalweighting of the two factors of RS

RSij =

[α(SKij )∑n

j �=i SKij

]+

[(1 − α)SBij∑n

j �=i SBij

]. (4)

Still building to the belief calculation, we require a value toquantify the expertise seeking (EXP) discussed at the beginningof this section. It is a pairwise function of the knowledge notshared between agents and their n alters and calculated using

EXPij =

∑k,i=j (KTMiik × KTMijk )|Ki | −

∑k KTMiik

(5)

We have noted that this model uses fact-based beliefs; there-fore, we had to provide the model a belief by fact weightingmatrix V that provides valence weights for each belief to eachfact from [−1.0,1.0]. This matrix allows facts to impact morethan one belief as well as have no impact at all with a weight ofzero (0). Intermediate outputs of Pythia—a different componentof the multimodeling effort that is a timed-influence Bayesiannetwork tool [38]–[40]—provided these weights and are shownin Fig. 2 as “Model & Experimental Params.” We will denotethe valence weight V for belief b using fact k as

Vbk . (6)

To account for the ability for an agent to have perceptions ofits beliefs, as well as to generalize self-perception to three (3)states (strongly agree, strongly disagree, and no opinion), weuse (7) for self-perception of belief b for later use

B′iib =

⎧⎪⎪⎨

⎪⎪⎩

1,∑

kVb k ×KTMi i k

|K | > 0.2

−1,∑

kVb k ×KTMi i k

|K | < 0.2

0, otherwise

⎫⎪⎪⎬

⎪⎪⎭. (7)

We also need to calculate the expected influence of self-perceived knowledge for each belief (EI), or how strongly theagent holds fact-based belief b, and we use the following equa-tion for that purpose:

EIiib =∑

k (Vbk × KTMiik )∑k Vbk

. (8)

We need three more values to calculate the per-agent per-belief value per-turn. They are the influentialness of alters onthe ego (INFj i) in (9); the resistance to being influenced byalters (INFii) in (10); and the total influentialness (TotalInfj i)in (11). The value of influentialnessj in (9) is an exogenousparameter set by the experimenter and represents the ability ofalters to influence the ego—the ji notation in (9) and (11) isintentional and not a typographic error. The value of BIi in (10)is also exogenous and represents the propensity of an ego tobe influenced by alters, also called belief influenceability. Inthis model, these values were kept constant and used successfuldefault settings from prior validated work with Construct.

∀j (alters) connected to i (ego)

INFj i = influentialnessj ×RSj i + EXPj i

2(9)

INFii = (1 − BIi) ×RSii + EIiib

2. (10)

106 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 44, NO. 1, JANUARY 2014

Summing (9) for each j connected to i helps us generate

∀j (alters) connected to i (ego)

TotalInfi =∑

j

INFj i . (11)

In compressed form, we can now represent agent i’s self-perception of belief b at time t in

BTMiibt= BTMiibt−1 × (1 − BIit

) + BIit

×

⎡

⎣

⎛

⎝∑

j

(INFj it

TotalInfit

× BTMijbt

)⎞

⎠

+(

INFj it

TotalInfit

× B′iibt

)]. (12)

V. INDIA–PAKISTAN CRISIS SCENARIO

We used this D2M approach, as operationalized in the soft-ware application AutoMap [28], [41], as part of a scenario-driven exercise. The intent of the exercise was to illustrate thevalue of two organizations coordinating to assess the impactof different courses of action (COAs). The organizations weretwo U.S. Regional Combatant Commands (COCOM): U.S. Pa-cific Command (USPACOM) and U.S. Central Command (US-CENTCOM). The scenario for this fictional situation used amixture of fictional scenario events and real-world events froma specific time period, from June 2, 2002 to August 5, 2002.It also used fictional and real-world interactions among agentsalong with real names for people and places. The interactionswere generated through SME elicitation and use five of theseven categories from [42]: public appeals; communication fa-cilitation; mediation; fact-finding; and humanitarian aid. Thescenario location is along the disputed territorial border regionsof Jammu and Kashmir between India, Pakistan, and China. Thescenario begins with a fictitious raid into the parliament buildingof Srinagar, India, by gunmen on June 2, 2002. The scenario con-tinues to August 5, 2002 with a number of actions by Pakistan,India, the U.S., and select other countries of interest.

A. Data-to-Model Process Applied to the Scenario

We used 3,000 LexisNexis-provided text files that met thesearch criteria of the scenario’s dates and the terms: “India” and“Pakistan.” These newspaper articles provided background andsupporting cultural and contextual data to the scenario-providedinformation. We also scraped each nation’s national securityapparatus’ official websites (circa 2010) as well as the offi-cial websites of USCENTCOM and USPACOM. By nationalsecurity apparatus, we mean the functional equivalents of theU.S. National Security Council, Department of Defense (DoD),and Department of State. After these web scrapes, there wereapproximately 27,000 files in our corpus. We built the syn-onym and classification thesauri as well as the delete list fromscratch: there was no COCOM planning staff from which wecould borrow thesauri or delete lists. The development of theselists took approximately 160 man-hours, although subsequent

TABLE IINODE COUNTS, PER COCOM, VIGNETTES A & B

improvements to Automap’s “Data to Model (D2M) Wizard”have demonstrated significant speedup [26].

Identification of specific persons relevant to a border-crisisscenario was an iterative process of identifying a term or setsof terms (e.g., “Prime Minister of India,” “Vajpayee”) and thenusing web-based searches to determine the nature of the termand resolve uncertainties. This allowed us to remove the multi-tude of cricket players and Bollywood stars within the corpus.We used social network measures, such as degree centrality,betweenness centrality, and eigenvector centrality, to estimatedifferent aspects of a node’s criticality in our resultant networks,as well as consultation with our multimodeling partners. Table IIdescribes the end state of the network model.

B. Network Analysis Applied to Scenario Models

Following the scenario outline, we divided the dataset intothree vignettes: 1) initial crisis incident plus eight days; 2) mid-crisis when the COCOMs were using independent analysis andactions; and 3) a final period when the two COCOMs would,in the scenario, collaborate and merge their respective modelsand COAs to present to U.S. national leadership. For each vi-gnette, we used the organizational risk analyzer (ORA) networkanalysis software [43] to calculate numerous static node andstatic network measures and to visualize the interconnectionsof strategic decision makers (labeled “SNA Reports” in Fig. 2).A more comprehensive discussion of which network measureswe used is available in Chapter 14 of the effort’s final technicalreport [44]. ORA includes over 157 different network measuresapplicable to two-mode and multimode networks [45].

Using this methodology, we discerned shifts in relative rank-ings of the top ten agents across the vignettes. Fig. 3 showsan example graphic from the “Key Agent” report, which is acomponent of the “Key Entities” report, for Vignette B from theUSCENTCOM perspective. The graphic identifies the agentsthat are most commonly in the top ten (10) rankings across 22different social network analytic measures relevant to agents. Inthis report, President Musharraf is in the top ten agents 90% ofthe time, or 20 of 22 measures. The numbers after the agents’title in the arrows reflect the change from Vignette A to VignetteB, with three new agents appearing in Vignette B. The Secretaryof State’s involvement suggests that the diplomacy instrumentof national power is increasing its level of effort.

The appearance of USPACOM and USCENTCOM, mean-while, is consistent with an interpretation that the scenariois rapidly moving from a diplomacy-centric situation to one

LANHAM et al.: SOCIAL NETWORK MODELING AND AGENT-BASED SIMULATION IN SUPPORT OF CRISIS DE-ESCALATION 107

Fig. 3. Change in actor relevance indicates that the scenario is shifting from adiplomatic to a military situation.

involving the U.S. military. This finding is in accordance withthe tenor of the scenario and the impressions of our SMEs. TheChairman of the Joint Chiefs’ (CJCS) drop in relative rank-ing is consistent with the increasing presence of both COCOMcommanders in direct discussions and interactions with the Pres-ident. Their direct involvement with the President is consistentwith the DoD moving from planning for action with the CJCSas the principal military advisor to executing action through theCOCOMs.

C. Dynamic Analysis Through Diffusion Simulations

The Automap-extracted metanetwork built by the D2M pro-cess, which was described in Sections V-A and V-B, became theprimary input to the Construct simulation, as shown in Fig. 2.For this scenario, the primary output measure of interest was thenumber of strategic decision makers who possessed a “pro-war”belief, as calculated using (12). We harvested the strategic deci-sion makers and their relationships directly from the text-minedmultimode data. The remainder of the multimode data, whichwere drawn primarily from LexisNexis, was not pertinent to thescenario and questions of interest.

We still needed some form of knowledge in the simulation;therefore, we implemented stylized representations of knowl-edge; these representations were created to reflect national iden-tity, culture, as well as knowledge both for and against aggres-sion. The size of these pools and their distribution to agents werebased on the text-mined outputs of AutoMap. We also seededthe simulation with the majority of agents to have more “anti-war” knowledge than “pro-war”—representing the status quothat there are far more people disinclined to go to war than thereare those inclined to start a war. The error-prone transactivememory was instantiated with a population-wide false-negativerate of 0.5 (that is, an ego wrongly perceives that an alter doesnot have a particular piece of knowledge).

We implemented the scenario within Construct as a set of41 exogenous “provocations,” 29 “responses,” and 24 generalevents that added general facts to agents’ knowledge pools.All these events had a magnitude, a start-time, and an end-time. Within Construct, we used a special-purpose agent foreach individual provocation and response. The knowledge bits

TABLE IIIEXPERIMENTAL CONDITIONS

these agents transferred were bit strings exogenously tied to thebeliefs of “pro-war” for provocations, and ‘anti-war’ for re-sponses. We modeled the impact of these events by constrain-ing the duration the special agents were active within thesimulation—inactive agents do not interact with others and,thus, do not share knowledge with other agents. The creation oflinks to decision-maker agents, as well as durations of activity,was drawn from the multimodeling team, SMEs, and interme-diate outputs of Pythia—all exogenous to the social networkfrom AutoMap. The duration of activity of the special agentsincreased the probability that the knowledge it was commu-nicating would be learned by its interaction partners, therebyimpacting the knowledge-anchored beliefs of those interactionpartners. The complete collection (∼120 kB) of provocationsand responses, durations, names, and start times, as well as thecomplete input files (∼400 kB) for Construct is available fromthe authors on request.

As shown in Table III, the principal question we exploredwas one of timing the interventions, with three relevant sub-questions, specifically: 1) how many strategic decision makerswill possess the pro-war belief if the U.S. does not intervene(the “None” case); 2) given the scenario and all of its deterrenceand de-escalation actions, how many Indian and Pakistani deci-sion makers will possess pro-war beliefs (the “Scenario” case);and 3) given a stable set of deterrence actions, how does chang-ing the timing of this action set change the number of decisionmakers with pro-war belief (the “Early,” “Middle,” and “Late”cases).

These virtual experiments showed that without U.S. or oth-ers’ work to tamp down tensions, within 30 days more than60% of the Pakistani and Indian strategic decision makers be-lieve that war is the right choice. Fig. 4 also indicates that theconventional studied diplomatic U.S. response set in the sce-nario document was insufficient to avoid war despite producinga shift toward antiwar beliefs in the minds of decision makers.Early interventions produced the most significant impact (seealso Fig. 5)—as agents then chose to pass along knowledge withnegative valence toward the “pro-war” belief.

108 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 44, NO. 1, JANUARY 2014

Fig. 4. Construct forecasts that the majority of strategic decision makers inboth India and Pakistan will possess the pro-war belief within 30 days. Deter-rence actions from the scenario have only minimal impacts until more than amonth after the crisis begins.

Fig. 5. Early interventions allow more time for comprehensive response.

There are a number of reasons for this outcome. Agents inthe model suffer the same “echo chamber” effect as people inthe real world—their interactions with agents like themselvesreinforce their beliefs and existing knowledge, forming a feed-back loop that gains even larger portions of the population [46].Interrupting that feedback loop early by exposing agents to ad-ditional or alternate information is critical. Fig. 5 was very sensi-tive to additional provocations (e.g., troop deployments, missilelaunches, riots, and media coverage thereof)—reinforcing theperception that actions will frequently overwhelm talks.

D. Model Implications for Policy

These results suggest that the U.S. and others must use leversof deterrence quickly; levers of deterrence may need repeateduse to have an effect; continued provocations will rapidly over-whelm U.S. instruments of national power; early and fast actionmay not, by itself, lead to de-escalation, although it may buytime to bring additional resources to bear.

E. Validation

Since the basis of the model is a fictitious scenario, thereis no mechanism available to do empirical or historical valida-tion of this specific model. As such, we turn to other forms ofconfirming face validity, plausibility, and usefulness of themodel. Of most important note, the conclusions from thismodel were consistent across three diverse sets of assumptions,paradigms, modeling languages, and tools: Construct, CAESARIII [47], [48], and Pythia. CAESAR III is a colored petri-net toolto assess decision-making organizations largely omitted fromthis discussion. The congruence in outcomes, despite the dis-tinctly different operating assumptions and paradigms of eachtool, provides increased confidence in the plausibility of the con-structed models, the techniques to build them, and the assess-ments that were derived from the master scenario event list. Thelarger multimodeling effort would also support incorporationof other models that include other motivations for interactions(e.g., social capital, exchange theory, and balance theory).

More important than the particular validation of this specificinstance is the confirmation that this semiautomatable andrepeatable approach of moving from large quantities of unstruc-tured text to a well-developed metanetwork is worthwhile. Thestatic analysis using social network analysis tools and techniquesgenerated reasonable results in the context of the scenario. Theapproach was further shown effective in generating the basisfor an agent-based simulation model that might otherwise havetaken significantly longer to build. Together, the presentedapproach shown in Fig. 2 provides techniques for decisionmakers to assess a wide variety of COAs in safe and controlledenvironments.

As for the scenario itself, the effort tells decision makers thatthere is little decision space within which they can maneuver—not a ground breaking result, but one based on more than intu-ition and the personal experience of individuals. This outcomewas also accepted by the SMEs of the project sponsor whenreviewing the effort.

VI. DISCUSSION

We demonstrated a rapid model development approach thatallows integration of multiple data sources to produce a metanet-work that forms the basis for the simulation. We showed the fea-sibility of how multiple organizations can take these metanet-works and examine possible futures—during long-range de-liberate planning and execution as well as crisis and time-sensitive environments. Finally, we established the ability ofdiffusion simulations and network science to provide estima-tions of action–reaction cycles in situations that require the co-ordinated use of multiple instruments of change—in this case,the elements of national power. These estimations were valid inthe eyes of the SMEs and the results were congruent with theoutput of models from very different backgrounds.

The first iteration of this approach and its application to deter-rence was a useful demonstration; however, there are challenges.Data drawn from LexisNexis are not an accurate reflection ofthe information each COCOM staff could or should maintain.We believe that as the data improve in quality and topicality,the utility and explanatory power of such models will improve.

LANHAM et al.: SOCIAL NETWORK MODELING AND AGENT-BASED SIMULATION IN SUPPORT OF CRISIS DE-ESCALATION 109

Furthermore, each COCOM’s information assets are likely tohave distinct and important differences that we did not reflect—and these differences may lead to diverse but more useful finalresults. Follow-on effort will need to incorporate a more sus-tained collection of text and other unstructured data. Fictionalscenarios will require additional synthetic data.

The diffusion simulation made some additional simplifyingassumptions. We did not conduct SME elicitation and profile thestrategic actor set to learn and program their starting inclinationstoward the pro-war belief. The simulation is able to use suchinformation, with a few minor changes. We relied primarily onthe agent-by-agent networks due to the paucity of the agent-to-knowledge links in the collected data. We estimate that COCOMstaff’s would have richer datasets that would support the use ofadditional networks within the Automap-extracted metanetworkand avoid or reduce the use of stylized knowledge sets withinConstruct. Another simplifying assumption was the deliberateexclusion of India’s Cold Start doctrine [49]–[51], their “no firstuse against nonnuclear states” policy [52], as well as Pakistan’spublished responses to the Indian doctrine. We did not incorpo-rate metacognition reasoning into the simulation—agents beingaware that others are attempting to influence them. Constructis very robust to trends and population/group level analysis. Itdoes not predict the precise actions of individuals at specifictimes nor should decision makers use Construct for per-agentanalysis or predictions.

VII. CONCLUSION

This study demonstrates that the D2M approach enables rapidmodel development and supports model reuse, merging, andextension when a network analytic approach is taken. This ap-proach meets the needs of decision makers to quickly model,simulate, and assess consequences of actions and reactions incrisis de-escalation environments.

ACKNOWLEDGMENT

The authors would like to thank their collaborators at GeorgeMason University’s System Architectures Lab. Without thecombined effort of A. Levis, Ph.D., S. A. K. Zaidi, Ph.D.,L. Wagenhals, Ph.D., R. Elder, Ph.D., T. Levitt, Ph.D., A. J.Y. Abu, and S. H. A. Rizvi, this project would not have metthe successes it did. This paper represents only a small part ofthe work this project entailed. They would also like to thank M.Kowalchuck for his expert knowledge.

REFERENCES

[1] J. Wilkenfeld, K. Young, V. Asal, and D. Quinn, “Mediating internationalcrises,” J. Conflict Resolution, vol. 47, pp. 279–301, Jun. 1, 2003.

[2] D. A. Sylvan and S. J. Thorson, “Ontologies, problem representation, andthe Cuban Missile crisis,” J. Conflict Resolution, vol. 36, pp. 709–732,Dec. 1, 1992.

[3] P. A. Schrodt and D. J. Gerner, “An event data analysis of third-partymediation in the Middle East and Balkans,” J. Conflict Resolution, vol. 48,pp. 310–330, Jun. 1, 2004.

[4] L. F. Richardson, Arms and Insecurity: A Mathematical Study of theCauses and Origins of War. Pittsburgh, PA: Boxwood Press, 1960.

[5] D. Ruloff, “The dynamics of conflict and cooperation between nations. Acomputer simulation and some results,” J. Peace Res., vol. 12, pp. 109–121, 1975.

[6] L. Yilmaz, T. I. Oren, and N. Ghasem-Aghaee, “Simulation-basedproblem-solving environments for conflict studies,” J. Simul. Gaming,vol. 37, pp. 534–556, Dec. 1, 2006.

[7] N. Eagle and A. S. Pentland, “Reality mining: Sensing complex socialsystems,” Pers. Ubiquitous Comput., vol. 10, pp. 255–268, 2006.

[8] M. Peter, “Ontologies are us: A unified model of social networks andsemantics,” J. Web Semant.: Sci. Services Agents World Wide Web, vol. 5,pp. 5–15, 2007.

[9] T. Berners-Lee, J. Hendler, and O. Lassila, “The semantic web,” Sci.Amer., vol. 284, pp. 34–43, May 17, 2001.

[10] N. Shadbolt, T. Berners-Lee, and W. Hall, “The semantic web revisited,”IEEE Intell. Syst., vol. 21, no. 3, pp. 96–101, May–Jun. 2006.

[11] H. Kautz, B. Selman, and M. Shah. (1997 Summer). The Hid-den Web, The AI Magazine [Online]. pp. 27–36. Available:http://www.cs.cornell.edu/selman/papers/pdf/97.aimag.refweb.pdf

[12] Y. Matsuo, J. Mori, M. Hamasaki, T. Nishimura, H. Takeda, K. Hasida,and M. Ishizuka, “POLYPHONET: An advanced social network extractionsystem from the Web,” J. Web Semant.: Sci. Services Agents World WideWeb, vol. 5, pp. 262–278, 2007.

[13] C. Weinstein, W. Campbell, B. Delaney, and G. O’Leary, “Modeling anddetection techniques for counter-terror social network analysis and intentrecognition,” in Proc. IEEE Aerospace Conf., 2009, pp. 1–16.

[14] A. McCallum, X. Wang, and N. Mohanty, “Joint group and topic discoveryfrom relations and text,” in Statistical Network Analysis: Models, Issuesand New Directions, Lecture Notes in Computer Science. vol. 4503, NewYork: Springer, 2007, pp. 28–44.

[15] J. Diesner and K. M. Carley, “A methodology for integrating networktheory and topic modeling and its application to innovation diffusion,”in Proc. IEEE 2nd Int. Conf. Social Comput., Minneapolis, MN, 2010,pp. 687–692.

[16] A. S. Rao and M. P. Georgeff, “Modeling rational agents within a BDI-architecture,” in Readings in Agents. San Francisco, CA: Morgan Kauf-mann, 1997, pp. 317–328.

[17] Z. Lin and K. M. Carley, “Organizational decision making and error in adynamic task environment,” J. Math. Sociol., vol. 22, pp. 125–149, 1997.

[18] M. Granovetter and R. Soong, “Threshold models of diffusion and collec-tive behavior,” J. Math. Sociol., vol. 9, pp. 165–179, 1983.

[19] N. Friedkin, A Structural Theory of Social Influence. New York: Cam-bridge Univ. Press, 1998.

[20] K. M. Carley, “An approach for relating social structure to cognitive struc-ture,” J. Math. Sociol., vol. 12, pp. 137–189, 1985.

[21] K. M. Carley, M. K. Martin, and B. R. Hirshman, “The etiology of socialchange,” Top. Cognitive Sci., vol. 1, pp. 621–650, Jun. 26, 2009.

[22] Y. Ren, K. M. Carley, and L. Argote, “The contingent effects of transactivememory: When is it more beneficial to know what others know?,” Manage.Sci., vol. 52, pp. 671–682, May 2006.

[23] K. M. Carley, “Group stability: A socio-cognitive approach,” in Advancesin Group Processes: Theory and Research, E. Lawler, B. Markovsky,C. Ridgeway, and H. Walker, Eds. Greenwich, CT: JAI Press, 1990,vol. III, pp. 1–44.

[24] C. Schreiber and K. M. Carley, “The impact of databases on knowledgetransfer: Simulation providing theory,” presented at the North Amer. As-soc. Comput. Social Organizational Sci., Pittsburgh, PA, 2003.

[25] K. M. Carley, “On the persistence of beliefs,” Sociological Abstracts,1991.

[26] K. M. Carley, M. Bigrigg, D. Garlan, J. Johnson, F. Kunkel, M. Lanham,M. Martin, G. Morgon, B. Schmerl, and T. van Holt, “Rapid ethnographicassessment: Data-to-model,” presented at the Human Social Cultural Be-havioral Focus: Integrating Social Sci. Theory Anal. Methods OperationalUse, Chantilly, VA, 2011.

[27] K. M. Carley, M. Bigrigg, D. Garlan, M. J. Lanham, Y. Lu, G. P. Morgan,and B. Schmerl, “Experimentation testbeds: Using SORASCS to run andprocess HSCB virtual experiments,” presented at the Human Social Cul-tural Behavioral Focus: Integrating Social Sci. Theory Anal. MethodsOperational Use, Chantilly, VA, 2011.

[28] K. M. Carley, D. Columbus, M. W. Bigrigg, and F. Kunkel.(2011, 4 Nov.). AutoMap User’s Guide 2011 [Online]. Available:http://www.casos.cs.cmu.edu/publications/papers/CMU-ISR-11-108.pdf

[29] J. Diesner and K. M. Carley, “Revealing social structure from texts: Meta-Matrix text analysis as a novel method for network text analysis,” in CausalMapping for Information Systems and Technology Research: Approaches,Advances, and Illustrations, V. K. Narayanan and D. J. Armstrong, Eds.Harrisburg, PA: Idea Group Publishing, 2005, pp. 81–108.

[30] K. M. Carley, “Dynamic network analysis,” in Dynamic Social Net-work Modeling and Analysis: Workshop Summary and Papers, R. Breiger,

110 IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS, VOL. 44, NO. 1, JANUARY 2014

K. M. Carley, and P. Pattison, Eds. Washington, DC: National AcademicPress, 2003, pp. 133–145.

[31] C. Schreiber and K. M. Carley, “Agent interactions in construct: An em-pirical validation using calibrated grounding,” presented at the Behav.Represent. Model. Simul. Conf., Norfolk, VA, 2007.

[32] C. Schreiber, S. Singh, and K. M. Carley, “Construct—A multi-agentnetwork model for the co-evolution of agents and socio-cultural environ-ments,” School of Computer Science, Institute for Software ResearchInternational, Carnegie Mellon University, Pittsburgh, PA, Tech. Rep.CMU-ISRI-04-109, Jan. 2004.

[33] M. McPherson, L. Lovin, and J. Cook, “Birds of a feather: Homophily insocial networks,” Annu. Rev. Sociol., vol. 27, pp. 415–444, 2001.

[34] C. E. Osgood, W. H. May, and M. S. Miron, Cross-Cultural Universals ofAffective Meaning. Urbana-Champaign, IL: Univ. Illinois Press, 1975.

[35] D. Krackhardt, “Social networks,” in Encyclopedia of Group Pro-cesses and Intergroup Relations, S. Otten, C. Cohrs, J. M. Levine, andM. A. Hogg, Eds. Los Angeles, CA: SAGE, 2010, vol. 2, pp. 817–821.

[36] B. R. Hirshman, J. St. Charles, and K. M. Carley, “Leaving us in tiers:Can homophily be used to generate tiering effects,” J. Comput. Math.Organization Theory, vol. 17, pp. 318–343, 2011.

[37] M. Kowalchuck, “Construct belief model,” Institute for Software ResearchInternational, School of Computer Science, Carnegie Mellon University,Pittsburgh, PA, 2012.

[38] P. W. Pachowicz, L. W. Wagenhals, J. Pham, and A. H. Levis, “Buildingand analyzing timed influence net models with internet-enabled pythia,”in Proc. SPIE Defense Secur. Symp., Orlando, FL, 2007, p. 292.

[39] L. W. Wagenhals, “Course of action development and evaluation usingdiscrete event system models of influence nets,” Ph.D. dissertation, Dept.Electr. Comput. Eng., George Mason University, Fairfax, VA, 2000.

[40] L. W. Wagenhals and A. H. Levis, “Course of action development andevaluation,” presented at the Command Control Res. Technol. Symp.,Fairfax, VA, 2002.

[41] J. Diesner and K. M. Carley, “AutoMap 1.2: Extract, analyze, represent,and compare mental models from texts,” Institute for Software ResearchInternational, School of Computer Science, Carnegie Mellon University,Pittsburgh, PA, Tech. Rep. CMU-ISRI-04-100, Dec. 2004.

[42] W. J. Dixon, “Third-party techniques for preventing conflict escalation andpromoting peaceful settlement,” Int. Organization, vol. 50, pp. 653–681,1996.

[43] K. Carley, J. Pfeffer, J. Reminga, J. Storrick, and D. Colum-bus. (2012, 5 Mar.). ORA User’s Guide 2012 [Online]. Available:http://www.casos.cs.cmu.edu/publications/papers/CMU-ISR-12-105.pdf

[44] K. M. Carley, I.-C. Moon, G. P. Morgan, and M. J. Lanham, “Adver-sary modeling—applications of dynamic network analysis,” in Compu-tational Modeling of Cultural Dimensions in Adversary Organizations,K. M. Carley and A. Levis, Eds. Fairfax, VA: The Volgenau School ofEngineering, George Mason University, 2010, pp. 172–203.

[45] K. M. Carley and J. Pfeffer, “Title,” Two-Mode and Multi-Mode Networks.Pittsburgh, PA: Institute of Software Research, 2012.

[46] K. Wallsten, “Political blogs and the bloggers who blog them: Is thepolitical blogosphere and echo chamber?,” presented at the AmericanPolitical Science Association Annu. Meet., Washington, DC, 2005.

[47] A. H. Levis and D. M. Perdu, “CAESAR II: A system for the designand evaluation of command and control organizations,” presented at the2nd Int. Command Control Res. Technol. Symp., Market Bosworth, U.K.,1996.

[48] S. K. Kansal, A. M. AbuSharekh, and A. H. Levis, “Computationally de-rived models of adversary organizations,” in Proc. IEEE Symp. Comput.Intell. Secur. Defense Appl., 2007, pp. 92–99.

[49] A. Ahmed. (2010, 4 Jan.). The ‘Cold Start and Stop’ strategy. Insitute forDefence Studies and Analyses—Comment [Electronic OpEd]. Available:http://www.idsa.in/idsacomments/TheColdStartandStopstrategy_aahmed_280910

[50] G. B. R. Kanwal. (2010, 4 Jan.). India’s cold start doctrine and strategicstability. Insitute for Defence Studies and Analyses—Comment [ElectronicOpEd]. Available: http://www.idsa.in/node/5442/372

[51] H. V. Pant. (2010, 4 Jan.). India’s quick-strike doctrine causesflutter, The Japan Times [Electronic Newspaper]. Available:http://search.japantimes.co.jp/cgi-bin/eo20100202a1.html

[52] S. S. Menon, “Speech by [Indian National Security Advisor] NSA ShriShivshankar Menon at NDC on “The Role of Force in Strategic Affairs,”National Security Council, Government of India, New Delhi, India, 2010.

Michael J. Lanham (M’11) received the M.S. de-gree in computer science (CS) from the University ofFlorida, Gainesville, in 2002. He is currently work-ing toward the Ph.D. degree with Carnegie MellonUniversity, Pittsburgh, PA under the Computation,Organization, and Society program.

He is a Lieutenant Colonel with the U.S. Armyand an Information Systems Management (FA53)Officer and former Infantry Officer. He has servedas U.S. Military Academy CS faculty, U.S. StrategicCommand/J65 Program Manager, Joint Functional

Component Command-Integrated Missile Defense/J35, J63, and Deputy ChiefInformation Officer, and U.S. Army Central Information Assurance ProgramManager. His current research interests include resilient command and controlin contested cyber environments.

Geoffrey P. Morgan received the B.S. degree in in-formation sciences and technology from the Pennsyl-vania State University, University Park, in 2005. Heis currently working toward the Ph.D. degree withCarnegie Mellon University, Pittsburgh, PA, underthe Computation, Organization, and Society program.

He has worked previously in industry, developinghigh fidelity models of human processes and proto-types of mixed-initiative robotic command and con-trol systems. His previous job titles include ArtificialIntelligence Engineer, Technical Lead, and Advanced

Technologies Lead. He has experience in agent-based modeling, cognitive mod-eling, and network-centric simulation. He is interested in organizations, indi-viduals, and how the two influence each other.

Mr. Morgan is an active member of the International Network for SocialNetwork Analysis and the Cognitive Science Society.

Kathleen M. Carley (M’06–SM’11) received theS.B. degree in political science and the S.B. degree ineconomics from the Massachusetts Institute of Tech-nology, Boston, MA, and the Ph.D. degree in mathe-matical sociology from Harvard University, Boston.

She is currently a Professor of Computation, Orga-nizations and Society Program with the DepartmentInstitute for Software Research, School of ComputerScience, Carnegie Mellon University (CMU), Pitts-burgh, PA, where she is also the Director of the Centerfor Computational Analysis of Social and Organiza-

tional Systems. She holds affiliated positions and courtesy appointments in theDepartments of Social and Decision Sciences, Engineering and Public Policy,and Electrical and Computer Engineering, CMU. Her research is focused ondynamic network analysis and agent-based modeling applied to issues such assecurity, counterterrorism, information diffusion, and social change.

Dr. Carley is a member of the International Network for Social NetworkAnalysis, American Statistical Association, Institute for Operations Researchand the Management Sciences, and the American Association for the Advance-ment of Science. She was the 2011 winner of the Simmel Award for outstandingcontributions to network science, and has served on numerous National Re-search Council committees and IEEE Workshop committees.