Noname manuscript No.(will be inserted by the editor)

SCAMP’s Stigmergic Model of Social Conflict

H. Van Dyke Parunak · Jason Greanya · Peggy McCarthy · JonathanA. Morell · Srikanth Nadella · Laura Sappelsa

Received: date / Accepted: date

Abstract SCAMP (Social Causality using Agentswith Multiple Perspectives) is one of four socialsimulators that generated socially realistic data forthe Ground Truth program. Unlike the other threesimulators, it is based on a computational principle,stigmergy, inspired by social insects. Using this ap-proach, we modeled conflict in a nation-state inspiredby the ongoing scenario in Syria. This paper summa-rizes stigmergy and describes the Conflict World webuilt in SCAMP.

Keywords Agent-Based Model · Social Simulation ·Psychology · Causality · Stigmergy

1 Introduction

The DARPA Ground Truth program used computersimulations (produced by four simulation teams) togenerate data from artificial societies. The two researchteams used this data to exercise and test their methodsfor extracting the underlying causality of the society.

H. Van Dyke Parunak · Jason GreanyaParallax Advanced Research, Beavercreek, OH 45431E-mail: van.parunak, jason.greanya@parallaxresearch.org

Jonathan A. Morell4.699... LLC, Ann Arbor, MIE-mail: jmorell@jamorell.com

Srikanth NadellaKairos Research, Dayton, OH 45458E-mail: srikanth.nadella@gmail.com

Laura Sappelsa · Peggy McCarthyANSER, Falls Church, VA 22041E-mail: laura.sappelsa, margaret.mccarthy@anser.org

To avoid leakage of causal information from thesimulation teams to the research teams, the two groupswere unknown to one another, and all data transfer wasby means of a test and evaluation (T&E) team [20].SCAMP (Social Causality using Agents with MultiplePerspectives) is one of the four simulators. 1 “MultiplePerspectives” in the acronym reflects SCAMP’s abilityto simulate different dimensions of causality.

The scientific contribution of this paper, as of theother simulator descriptions in this special issue, is todocument how the features of the simulation supportthe experiments in the overall program. Other papers[25,27,26] and an ODD protocol [14] for SCAMP [24]provide technical details on SCAMP itself, though Sec-tion 2 provides a brief overview of SCAMP’s distinctivefeatures, particularly in comparison with the other sim-ulators in the program.

The program ran through three challenges, with suc-cessively more complex models in each challenge. Mostof the data in this paper comes from the third challenge.Challenges 1 and 2 supported the event (Section 5) andgoal (Section 6) perpectives, while Challenge 3 addedthe geospatial (Section 7) and social (Section 8) per-spectives. In each challenge, the research teams facedthree tests: explain the underlying ground truth on thebasis of an initial data package containing simulationresults over an extended period, predict how the sce-nario would evolve in the future based on this groundtruth, and under different proposed changes, and pre-scribe changes to achieve certain objectives.

Section 2 outlines SCAMP’s distinctive architec-ture, introducing stigmergy and briefly comparing ourapproach with other simulation technologies in general

1 The authors are grateful for the consulting support ofKathleen Carley of CMU and Netanomics on social networkissues.

2 Parunak et al.

and the other simulators in Ground Truth in particular.Section 3 outlines the scenario in which we embeddedour simulation. Section 4 describes the different groupsinvolved in the conflict, and provides more detail onarchitectural features that support group distinctionsamong agents. The next four sections describe thefour perspectives that SCAMP currently supports:the types of events in which these agents can partic-ipate (Section 5), the goals that each group pursues(Section 6), spatial constraints (Section 7), and socialdynamics (Section 8). Each of these four sectionsbegins with the mechanics of the perspective and thetools used to model it, then presents some sampledata to show how that perspective contributes toSCAMP’s social and psychological realism. Section 9discusses the Predict and Prescribe tests that we gavethe research teams, and Section 10 concludes.

2 SCAMP and Other Social Simulators

In this section, we briefly review the very large spaceof technologies used for social simulation. Then we de-scribe the distinctive features of SCAMP, and compareit with the other simulators used in the Ground Truthprogram.

A wide range of technologies have been used forsocial simulation [6,19,9,12]. The major distinctionamong them is between equation-based approachessuch as system dynamics [43] and its qualitativefoundation of causal loop diagrams, and agent-basedapproaches. Equation-based simulations track theevolution of measurable variables during a system’soperation, typically using differential equations, andtheir structure makes it convenient for them to focuson population averages over the variables describingindividual actors, using what a physicist would terma mean-field approach [23]. Agent-based approachesnaturally represent each actor’s variables separately,and if the dynamics of the simulation are complex,the values of these variables can diverge from oneanother, yielding an overall result that is qualitativelydifferent from an equation-based model of the samescenario [35]. All four simulators used in Ground Truthwere agent-based. While equation-based simulationsrun faster than agent-based ones, advances in com-putational technology have led to increasing use ofagent-based models, and several frameworks supporttheir development, including Repast [1], MASON [18],and NetLogo [49].

The fundamental property of an agent is that it con-tinuously monitors its inputs, reasons about them, andtakes action. Various agent-based simulators differ inthe kind of internal logic used by the agents in reasoning

from inputs to outputs. Common approaches includearchitectures that explicitly model the beliefs, desires,and intentions of individual agents (BDI agents, [38]),Bayesian reasoning (often applied in a BDI framework[5]), and raw conditional logic driven by numerical statevariables.

SCAMP is based on an agent architecture [21,30,31,34] known as “stigmergy.” Grassé [13] coined this wordin 1959 from the Greek στίγμα (sign) and ἒργον (action)to describe insect actions that are mediated by signs inthe environment, such as pheromones, rather than bydirect inter-agent messages.

Fig. 1 shows schematically the interaction betweenagent and environment. An agent’s state and local envi-ronment determine its actions, and are modified by itsactions. The environment modifies its own state, andagents interact by sensing signs left by other agents.

Because of our ignorance of linguistic capabilitiesin non-human organisms, stigmergy is widely used tomodel animal communities [4,2,45]. Agents are local-ized in a graph-structured environment (for ecologicalstudies, typically a geospatial lattice). Each time step,they “deposit pheromone” by augmenting some vari-able associated with their current location, and sensethe value of this variable in adjacent locations to guidemovement decisions.

The relevance of stigmergy to human behaviordraws on an observation by Herbert Simon [41] that isknown as “Simon’s Law.” He writes: “An ant, viewed asa behaving system, is quite simple. The apparent com-plexity of its behavior over time is largely a reflectionof the complexity of the environment in which it findsitself.” Then he extends this insight to human behavior:“Human beings, viewed as behaving systems, are quitesimple. The apparent complexity of our behavior overtime is largely a reflection of the complexity of theenvironment in which we find ourselves.” The trickis to encode the agents’ complex behavior, not inthe agent code, but in the environment in which theagents live. While conventional cognitive agents putthe domain model inside the agents, SCAMP putsthe agents inside the model. That is, the agents moveover external graphs reflecting the choices availableto them, including both spatial lattices (Section 7)

Fig. 1 Basic stigmergic schema

SCAMP’s Stigmergic Model of Social Conflict 3

and directed graphs (Section 5)). Different kinds ofagents represent different geopolitically active groupsto which they belong (Section 3). Each node in thesegraphs has a series of presence variables for each kindof agent. These presence variables support a digitalimitation of insect pheromones. Agents augment thepresence variables of their own kind for the nodes theyvisit, and sense the values of all presence variables inadjacent nodes when considering a move. These graphsand the variables that they support constitute theenvironment invoked by Simon’s law, and determinethe behavior of the agents.

In a purely stigmergic system, agents interact onlywith their local environment. SCAMP also allowsagents to exchange information directly with agentswhom they meet on a node of the environment (Sec-tion 8, but this exchange is highly stereotyped. Thus,unlike the other Ground Truth simulators [37,50,36],SCAMP does not support remote messages betweenagents. However, it does support two forms of informa-tion movement among nodes of the environment thatcan allow one agent to react to the actions of otheragents that are remote from it: influence edges in theCausal Event Graph (Section 5), and changes in theurgency of one event based on other remote events byway of the goal system (Section 6).

A SCAMP agent repeatedly chooses among ac-cessible alternatives. The basic alternatives are eventtypes in a directed graph (Section 5) and locations ina geospatial lattice (Section 7). An agent is alwaysparticipating in one event type, and considering thenext. It always has a current location, and some eventtypes require it to move in geospace toward a goalspecified in the definition of the event type.

Every alternative open to an agent (an event type,or a geospatial location) is a node in a graph, and ischaracterized by a vector of scalar features (Section 4).These include the presence variables mentioned above.Each agent has a vector of scalar preferences over thissame feature space, based on the groups with which itis affiliated. To choose among alternatives, it computesthe cosine (normalized dot product) between its pref-erences and the features of each alternative, exponenti-ates each value (to make them positive), and normalizesthe set to form a roulette wheel that it spins.

Each entity in the model is represented by a polya-gent [29], a single avatar that continuously deploysa swarm of ghosts to explore the future. The ghostsexecute the selection logic described above. As theymove, they deposit pheromone by augmenting modifythe presence features of the event types or geospatiallocations they visit (the arrow in Fig. 1 from AgentDynamics to Environment’s State.

In each decision cycle, the avatar sends out a lim-ited number of waves of ghosts, each consisting of a lim-ited number of ghosts. Each ghost carries its avatar’spreference vector. Each ghost explores the future to alimited horizon. Then the avatar follows the crest ofthe presence pheromones deposited by its ghosts. Inpsychological terms, the ghosts implement the commonhuman practice of planning by mental simulation [16].Each ghost represents an actor’s mental projection ofone possible future, and repeated waves of ghosts reflectan actor’s repeated mental simulation informed by theresults of a previous cycle (“Hmm, that didn’t turn outwell. Let’s think this through again ....”). The limitson number of waves, number of ghosts, and explorationhorizon are important bounds [42] on SCAMP’s ratio-nality.

As a ghost from one group runs into the future, itmay encounter features modified by ghosts from othergroups, also in the future. If its preferences are sensitiveto those features, it will respond to them, with a degreeof attention modulated by its preference vector (Sec-tion 4). At first glance, the notion that the ghosts be-longing to one avatar deposit presence features that canbe sensed by other avatars seems to violate the conceptof ghosts as an individual avatar’s mental simulation ofpossible futures.2 In fact, this facility gives our avatars aprimitive theory of mind that allows them to reason re-cursively about the reasoning of other groups. A ghostcannot see other ghosts, only the local value of theirpheromones (their presence features). This pheromonefield corresponds to a probabilistic estimate of wherethose ghosts actually were and where their avatar maymove. While our implementation generates this esti-mate based on the actions of other agents, the nature ofthe information available to an avatar’s ghosts is con-sistent with that produced by a local theory of mindinformed by observations of other agents.

This decision process, based on the mathematicalmanipulation of the features of decision alternativesand the preferences of each agent, is mechanical. Stig-mergic behavior by people is well documented [22], butstigmergy might seem incapable of replicating the kindof higher-level biases that dominate human thought.In fact, the processing provided by each of SCAMP’sperspectives opens the door to behaviors that can bealigned with many recognized cognitive processes andbiases [25].

While not commonly used for social modeling, stig-mergy offers a number of benefits over more conven-tional architectures.

2 We are grateful to a referee for requesting this clarification

4 Parunak et al.

– Speed of processing: The computations involved (dotproducts and roulette selection) are extremely effi-cient on modern processors.

– Extensibility: Additional perspectives can easily beadded to the model, as long as they meet one ofthree conditions. 1) They can be traversable graphs,that is, graphs over which it makes sense for agentsto move by comparing agent preferences with nodefeatures (like our current event graph (Section 5)and geospatial (Section 7) perspectives). 2) Theycan take as inputs the features of nodes on an ex-isting traversable graph and deliver their output aschanges to such features, as does our current goalsperspective (Section 6). 3) They can modulate thepreferences of agents, as does our current social per-spective (Section 8).

– Modeling alternative futures concurrently: The mul-tiple ghosts sent out by an avatar explore alterna-tive possible futures for the avatar in a single run,and the pheromone field that ghosts develop over atraversable graph is, up to a normalizing constant,isomorphic to a probability field giving the proba-bility that the corresponding avatar will be at eachlocation at a given time. Currently, we use this fieldonly to guide the avatar, but it could be analyzedto provide information on the likelihood of differ-ent possible outcomes, without the need for costlyrepetitive simulations.

– Model accessibility: The logic of SCAMP’s mode isexternal to the individual agents, encoded not incomputer code, but in artifacts accessible to ana-lysts who are not programmers. Modelers constructthese artifacts using CMapTools [15] for networkstructures, GIMP [46] for geospatial information,and Excel for detailed parameters. This inversionof the relation between agent and domain modelmakes the behavior of SCAMP much more acces-sible to domain experts who are not programmersthan traditional approaches.

One disadvantage of a purely stigmergic model isthat agents cannot exchange messages directly with oneanother. SCAMP relaxes this constraint in the socialperspective (Section 8): if two agents participate in thesame event type or visit the same geospatial tile con-currently, they exchange their preference vectors anduse them (according to individual parameters) to mod-ify their own preferences, modeling the influence of ourassociates on our own attitudes. Though highly stereo-typed, this extension demonstrates that a stigmergiccode base can be extended to provide more conventionalfeatures, such as a rich interagent language.

It is illuminating to compare SCAMP with the otherthree social simulations in the Ground Truth program,

each using a different agent-based social modeling tech-nology. Each of these systems is discussed in further de-tail elsewhere in this special issue. We highlight in thiscomparison the ability of non-programmers to createand modify the models [27].

George Mason University, Tulane University, andthe University of Buffalo produced a model of UrbanLife [50] in the MASON modeling toolkit [18] andits GeoMASON extension [44]. Numerous aspects ofthe model are generated algorithmically, including theagent population, the geospatial map, and the socialnetwork among the agents. The system provides apredefined set of triggers (sensitive to both internal andexternal factors), behaviors to which they lead, actionsthat make up behaviors, and goals that determine whenactions stop. Defining new triggers, behaviors, actions,and goals requires programming, but a drag-and-dropinterface allows modelers to assemble and parameterizethese components to define a scenario.

Raytheon BBN produced ACCESS [37] in theRepast framework. The model highlights the inter-actions among individual agents, groups to whichthey belong, and the overall population or “world.”ACCESS models space as a list of locations, butwithout orientation or distances, so there is no “map”for a user to enter, and the individual behaviors aredetermined by equations embedded in the code.

USC ISI produced a disaster world [36] in theirPsychSim social simulation framework. Agents aredriven by partially observable Markov decision pro-cesses (POMDPs) and can reason recursively aboutone another. PsychSim provides one interface thatallows social scientists to create simulation modelsdirectly, and another allowing them to manipulatethe parameters governing the simulation. However,both interfaces abstract over the full complexity of aPsychSim model (e.g., limiting the types of probabilitydistributions and reward functions), so specifying arbi-trary probability and utility models requires sufficientprogramming ability to use the PsychSim API.

GMU and ISI both support non-programmers whowish to modify a scenario, but still require programmersto modify details. In addition, they introduce propri-etary interfaces. SCAMP allows non-programmers todefine new groups, actions, and goals and their rela-tions, using tools with which they may already be fa-miliar.

3 Scenario

The model we built in SCAMP reflects multipartisanconflict in a country inspired by (but not a replica

SCAMP’s Stigmergic Model of Social Conflict 5

of) Syria, with an authoritarian government seeking tocling to power in the face of a democratic armed op-position and an ideological extremist force seeking toestablish a religious state. Relief agencies seek to re-lieve the condition of the general population, who justwant to get on with their lives. Section 4 defines thesegroups in more detail. Over the course of the programwe modeled two imaginary countries, Donglap and Tha-rum.

A particular challenge in the Ground Truth programwas enabling the research teams to interact with thesimulation teams as though they were interacting witha real social scenario, rather than with computer scien-tists seated at terminals. To help us structure a coher-ent interaction with the research teams, the SCAMPteam adopted the persona of the Fourth Marine Ex-peditionary Battalion (MEB), tasked with a variety ofoversight and stabilization assignments in Donglap andlater in Tharum. We cast the research teams as advisorsto our civil affairs officer. We sought to stay in characterwith this persona in all interactions with the researchteams. In particular, our tasking memos were deliveredas though they were prepared on a manual typewriter(we were, after all, deployed to remote locations), sub-ject to appropriate military approvals (Fig. 2). Limi-tations to research requests were explained in terms ofrealistic operational constraints on a military unit. Pre-dictions envisioned events that the Fourth MEB sus-pects might occur or actions that it might reasonablytake in line with its mission and general military policysuch as rules of engagement and status of forces agree-ments, and prescriptions were motivated as guidance tothe Fourth MEB for future actions.

Here is the tasking memo for the Explain test inChallenges 1 and 2 (which used the same ground truth).We comment on the rationale for the various details andthe constraints they pose on our model.

You are advising the civil affairs officer attachedto the 4rd Marine Expeditionary Brigade, re-cently deployed to Donglap, a new country that

Fig. 2 Example Tasking Letter

formed after the dissolution of an unstable thirdworld country.

This role is a realistic one in view of DARPA’s mission,and motivates the interaction of the research teamswith us. However, it does make them an extendedpart of our organization. To avoid any questions aboutwhether their interaction with us might influencethe scenario, we did not model the Fourth MEB asa causal player in our ground truth. However, someof the changes that we asked the research teams toexplore in the Predict and Prescribe tests were actionsthat the Fourth MEB could take to modify the groundtruth (for example, refuse access to a geospatial region,suppress events of a given type, or modify the numberof actors associated with a given group, whetherthrough diplomatic channels or direct military action).

Donglap is in the midst of a civil conflict notentirely unlike the current Syrian situation. OurCO has tasked you with figuring out how thesociety works, so that US operators can winthe support of the local population rather thanalienating them through unintentional gaffs.The locals are understandably nervous abouttalking with outsiders, but we have found oneactor whom you can interview, and from whomyou can learn more. We will identify actorswith labels like A8263. In this culture, extendedinterrogations are considered rude, and any oneactor will tend to stop cooperating if you queryit too much.

SCAMP runs offline, recording its results to logs fromwhich we respond to queries. Early in the program, wewere concerned that the research teams might requestall our data at the outset, leaving us with no more togive them. A shy populace and cultural taboos againstextended interrogations provided a valve to control theflow of information. The lack of data turned out to bea problem, and in later challenges we generated longerlogs and were more generous with our data.

Actors participate in different events over time.We will provide you with succinct descriptionsof some of them, but most you can recognizeonly as distinct events, identified with labels likeE32. There are two reasons for this ambiguity:your unfamiliarity with the culture, and the re-luctance of the actors to talk to strangers aboutsome of their activities. (This is, after all, a timeof war, and one sometimes doesn’t know the loy-alty of the person to whom one is speaking.)However, over time, as you learn more about theculture, the meanings of some of these events

6 Parunak et al.

may also be revealed. We will also provide youwith a history of environmental events, that is,events caused by forces other than the actors wesimulate.

Our events have very full descriptions (Section 5), andDARPA was concerned that the research teams wouldbecome so preoccupied with the causal implications ofthe descriptions themselves that they would not focuson extracting causality from the behavioral data weproduced. This part of the scenario justifies the obfus-cation we applied to event names to avoid this problem.

Our persona also supported contextualization of ex-periments requested by the research teams. Direct ex-periments with subsets of the population (e.g., focusgroups) seemed out of keeping with our persona, butwe adopted the following mechanism:

Unfortunately, there are many conflicts goingon in the world today. Many involve similarcultures to the one we are presenting you (sothat the causal structure remains the same,though details may differ). As you discoverour ground truth, you can request data fromanother conflict that differs from our baseconflict in specified elements of the groundtruth. To make such a request, you must learnenough about our ground truth to specify whatelements you want changed. It is likely thatthe other conflict will also differ in other waysbeyond your control and knowledge, but it willhave the same causal structure as the centralconflict you have been studying.

This rationale is not only supports experimentation,but also allows us an excuse for rejecting experimen-tal requests that come too close to asking us to solvePredict and Prescribe tests that we posed to the re-search teams. We simply replied, “We can’t find sucha country.” Sometimes we justified this with the addi-tional note, “If we were able to find such an example,we wouldn’t need to ask you for the prediction” (respec-tively, prescription).

In Challenge 3, we added two new perspectives toour model: geospace (Section 7), and social dynamics(Section 8). These changes significantly enlarged theunderlying causal structure. To help the research teamsavoid confounding the new structure with the old one,we redeployed the Fourth MEB from Donglap to Tha-rum. Here is the tasking memo for Challenge 3.

The 4rd Marine Expeditionary Brigade has beenredeployed to Tharum. Like Donglap, Tharum issuffering from internal unrest that is stimulatingrefugee activity, but it differs from Donglap inthree main ways.

1. We can now provide you with a geospatialmap of the terrain as well as the current loca-tion of actors over time. Actually, we providetwo maps: elevation data (light is higher),and one with features such as roads, water,locations of cities, and boundaries of coun-tries. Tharum’s culture is rooted in ancientBabylonian society, whose mathematics werehexadecimal, and so location is measured ona hex grid, with (0,0) at the upper left cor-ner and (23, 20) near the lower right (the firstnumber indicates horizontal position; the sec-ond, vertical). In reporting the history of ac-tors, we will give you not only the sequenceof events in which they participate and theirsatisfaction level, but also their geospatial co-ordinates. Tharum is small enough that en-vironmental events (“nature”) apply to thewhole country, and we do not give coordi-nates for them.

2. Most of the event types that we experiencedin Donglap also occur in Tharum, but wehave identified about 200 more, and there isno guarantee that the relation of the onesthat you’ve seen before to one another is thesame as it was in Donglap.

3. People in Tharum are more open about theirassociations than they were in Donglap, andwe are now able at each point in time to tellyou, for each actor, what other actors it hasencountered in its history up to this point,and how close that relationship is.

For computational reasons, we use a hex grid for geo-space, and the first item creates a backstory for thisunusual reference system, as well as introducing thenew geospatial data that is available. The third itemsupports the new social perspective.

As before, we are trying to collect data on con-flicts elsewhere that share the underlying causalstructure of Donglap, and on request, will try tofind data for you from other conflicts that differsfrom our base conflict in specified elements of theground truth. To make such a request, you mustspecify what elements you want changed. Youalready know that it is sometimes possible tofind situations that do not include certain eventtypes, or in which the relative populations of dif-ferent types of actors differ. Some cases in ourarchive also include situations in which a spe-cific event becomes inaccessible partway throughthe data. It is likely that the other conflict willalso differ in other ways beyond your control and

SCAMP’s Stigmergic Model of Social Conflict 7

knowledge, but except for the changes you re-quest, it will have the same causal structure asthe conflict in Tharum.

In Section 9, we will show how we extend this per-sona to posing Predict and Prescribe problems for theresearch teams.

4 Groups and Feature Space

Feature space is the vector space in which event fea-tures and agent preferences are defined. The structureof this space depends on the distinct groups of agentsin a model. Every agent is affiliated with one or moregroups. Conflict World has six groups representing dif-ferent agents in the conflict:

1. The government (GO)is authoritarian, bent on re-taining its own control of the situation, and willingto oppress its people to keep them in line.

2. The military (MIL) is initially aligned with the gov-ernment, but can diverge. (Challenges 1 and 2 didnot have a separate MIL group.)

3. The armed opposition (AO), inspired by the Syrianopposition, is a movement from within the countrythat seeks to reform or replace the government withdemocratic institutions. Our mission as the FourthMEB, reflected in our Predict and Prescribe tests,is to promote this change.

4. The violent extremists (VE), inspired by ISIS, are anideologically driven foreign faction that seeks to in-clude the local territory (whether Donglap in Chal-lenges 1 and 2 or Tharum in Challenge 3) in a largerreligious state.

5. Relief agencies (RA) seek to provide humanitarianrelief for civilians, largely in the form of refugeecamps both within and just outside of Tharum.

6. People (PEO) are pro-opposition, anti-governmentcivilians, just trying to get on with their lives.

Each group has a hierarchical goal network (HGN)described in Section 5, and a vector of preferences in[-1, 1] over the same feature space that encodes thefeatures of alternatives that agents may consider. Thedimensions of feature space are of three types: exoge-nous, urgency, and presence.

– Exogenous features (each in [-1, 1]) are defined bythe modeler, and in Conflict World are three innumber: economic, physical, and psychological well-being. As event features, these describe the impactof the event type on the agent, while as preferences,they describe the agent’s priorities. A positive fea-ture means that the event type improves that facet

of agents who participate in it (for example, an eventtype “go to work for the day” would have positiveimpact on a participating agent’s economic well-being, reflecting the expectation of income), whilea negative feature indicates that an event type willreduce an agent’s well-being (“participate in streetriot” would have a negative physical feature, re-flecting the risk of physical harm). These featurescorrespond to positions in each agent’s preferencevector. For instance, a preference of 0.1 for physi-cal well-being means that agents tend not to caremuch about their health and safety, while a prefer-ence of 0.9 means that agents are very much con-cerned about this feature. While the decision mech-anisms can support masochistic agents (with well-being preferences < 0), we restricted preferences forexogenous features to values > 0 for the GroundTruth program.

– Each group has one urgency feature in feature space,reflecting the current state of that group’s HGN.One group’s preference for another group’s urgencyfeature reflects its desire to promote (preference >0) or block (preference < 0) the other group’s goals.

– Each group has one presence feature. In imitationof insect pheromones, a group’s feature on an eventtype or location is augmented each time a ghost rep-resenting the group visits that alternative (the ar-row from Agent Dynamics to Environment’s Statein Fig. 1), and evaporates over time (Environment’sDynamics in Fig. 1). One group’s preference for an-other group’s presence feature reflects its attractionto or repulsion from members of that group. Evap-oration of the presence feature imposes a tempo-ral limitation on knowledge of past participation ofagents in events or locations. It is an important ex-ample of the bounded rationality [42] of SCAMPagents, and an instance of recency bias.

As noted in Section 2, a ghost responds to the pres-ence of ghosts from other groups if its preference forthe presence features of those groups is non-zero. Sinceghosts are exploring possible futures for their respectiveavatars, the strength of the presence features for a givengroup reflects the probability that an avatar of thatgroup will be at that location at that time, and whenghosts attend to the presence features of other ghosts,they are behaving recursively [47], reasoning about thefutures being considered by the avatars represented bythose ghosts. As in other recursive agent formalisms,the recursion is nested: agent A reasons about whatagent B thinks about what agent A thinks about what.... The depth of the recursion is the number of wavesof ghosts that an avatar sends out, and the impact ofthe recursion depends on the ghost’s preference for the

8 Parunak et al.

Table 1 Key group parameters in Conflict World

Group Population PreferenceVariation

AffiliationThreshold

Government 13 0.2 0.98

Military 6 0.2 0.98

ArmedOpposition 11 0.3 0.98

ViolentExtremists 8 0.3 0.98

ReliefAgencies 4 0.3 0.90

People 18 0.4 0.80

presence feature of the other group. Both of these pa-rameters depend on the ghost’s group.

The preference vector for each group is a baselinefrom which the preferences of individual agents are gen-erated. Each group is assigned a number of initial agentsand a variation v in [0, 1]. Each preference for a newagent is selected uniformly from the baseline value ± v.

Once an agent has preferences, it considers whetherto affiliate with other groups, based on the cosine ofthe angle in feature space between its preference vectorand the baseline preference vectors of the other groups.Each group has a threshold that the cosine must ex-ceed for affiliation to take place, and the cosine definesthe weight of the affiliation. The preference vector thatthe agent uses in choosing alternatives is the weightedaverage of its own preference vector and those of othergroups with which it is affiliated, and its ghosts aug-ment presence features of all affiliated groups, againproportional to the affiliation weights. In both cases,the weight for its home group is 1, and the weights forthe other groups are its cosine proximity to them. Itspreferences for exogenous features vary with its experi-ence, while its preferences for urgency and presence fea-tures are fixed unless the social perspective (Section 8)is active.

Table 1 shows each group’s initial populations, pref-erence variations, and affiliation thresholds. Note therelative homogeneity of the Government and Military(low variation) compared with other groups, and thewillingness of People to affiliate with other groups.

At first glance, the relatively small initial popula-tion (60 agents) may seem unrealistic. However, eachavatar decision comes from a swarm of 48 ghost agents,so the actual exploration of alternatives involves nearly3000 agents. In addition, the social dynamics perspec-tive (Section 8) allows new agents to join the simulationas it runs (for example, by influx of foreign fighters). In

our baseline run for Tharum, the simulation ended with548 avatars, represented by more than 26k ghosts.

Each group also specifies one or more geospatial re-gions over which its agents are initially distributed.

In some configurations, we add Neutral agents,who have no home group and no baseline preferences.Their preferences are assigned randomly, and thenwith threshold 0 they affiliate with the closest group.Neutral agents are not necessary if we can estimate inadvance the proportion of actors in each defined group,but they allow us to model a population with a largeproportion of actors of unknown group affiliation. Inthe Ground Truth program, they also provide a leverto adjust the complexity of the simulation.

SCAMP also supports an Environment group whosesingle agent moves over a subgraph of backgroundevents to generate natural events (droughts, famine),or events who detailed causality is not modeled in thesystem (economic downturn, assistance from foreigngovernments). The Fourth Marine ExpeditionaryBattalion is not a group in the Conflict World, andparticipates in the causal dynamics only by modifyingthe ground truth for Predict or Prescribe tests.

Groups are defined in a worksheet in the model’sExcel workbook.

5 Event Types

The fundamental dynamic of SCAMP is agent choiceover alternatives, and the main alternatives arerecorded in a Causal Event Graph (CEG), a directedgraph whose nodes are event types. Every trajectorythrough this graph is a coherent narrative for anagent’s experience.

5.1 Event Types and the Causal Event Graph

There are over 400 event types in Tharum. These in-clude:

– large numbers of people move to urban areas– public demands democratic reforms– Government security forces arrest minority leader– military refuses to carry out government’s orders– govt & opposition leaders commence official talks– Neighbors leave– people arm themselves– relief agencies identify an increase in unplanned

need– Military bombs opposition-controlled neighbor-

hoods– funders of relief agencies lose interest in conflict

SCAMP’s Stigmergic Model of Social Conflict 9

– Protesters share political news on social media– QOL at IDP camp improves– government & armed opposition forces cease nego-

tiations– head of state/government calls for end to violence– transitional government invites election monitors

We call these “event types” because they can recurover the course of a run. Strictly speaking, an “event”is a consecutive period of time during which there areagents participating in a given event type. However,where there is no risk of confusion, we often refer tothe nodes in the CEG as “events.”

Each event type has a feature vector over the samefeature space that defines group parameters. The ex-ogenous features are coded by the modelers, while theurgency features are modulated by the current state ofthe world via the hierarchical goal networks (Section 5)and the presence features are augmented each time anagent participates in an event, and evaporate exponen-tially through time. For example, consider a stationaryagent that augments its group’s presence feature at itscurrent location by d each time step, while the featureevaporates each step by the factor e ∈ (0, 1). The mostrecent deposit contributes d, the one from the previoustime step de, the one from two time steps back de2, andso forth. Thus the total presence feature after n stepsis d+ de+ de2 + ...+ den−1 = d

∑n−1i=0 ei, which is just

the geometric series, and asymptotes to d/(1− e).Not every event type makes sense for all groups.

For example, only Government agents can meaningfullyparticipate in “Government security forces arrest minor-ity leader.” Each event type is scored to indicate whichgroups have agency for it. Some of the directed edgesthrough this graph are agency edges, connecting eventtypes for which the same group has agency, and indicat-ing that an agent currently participating in the originof the edge may choose to traverse the edge to reach itsnext activity.

The nodes and agency edges together form a narra-tive space [39]. Any single trajectory through this spaceis a plausible narrative for agents that follow it. Forcomputational convenience, the narrative space beginswith the node START and ends with the node STOP,so that every valid narrative is a path from START toSTOP.3 To generate trajectories longer than those en-coded in the CEG, we send agents that reach STOPback to START. Since their own preferences and thefeatures of the events have changed, they may very wellfollow a different trajectory on each transit of the CEG.

3 In the CEG for Ground Truth, each group has its ownSTOP node, but to simplify analysis we envision an additionalSTOP node with an incoming edge leading from each of these.

For example, here is the trajectory of an agent be-longing to the People group:

– people desire a government that recognizes civil lib-erties and legitimizes all sectors of the population

– civilians are dissatisfied with the authoritarian govt– public demands democratic reforms– protesters attack government facilities & destroy as-

sets– government and protesters clash– Protesters attack police station killing officer and

burning it down– large numbers of people protest throughout the

country

Each agent maintains a memory of its entire life his-tory, and each time an agent moves, it logs its movementfor later analysis.

An important part of intelligence analysis (or sce-nario modeling) is anticipating possible patterns of be-havior that the actors of interest might exhibit. Ana-lysts commonly describe scenarios in terms of possiblenarratives, making the CEG a natural representationfor capturing complex social situations, and the CEGwas originally developed in support of intelligence anal-ysis. One of the benefits of this representation is that itamplifies the creativity of analysts by combining narra-tives that they explicitly formulate to yield a huge num-ber of other narratives that are consistent with these.Ground Truth is not concerned with analytic creativity,but this same amplification means that the CEG cangenerate an incredibly large number of different behav-ioral trajectories as data for the research teams.

A simple example illustrates this amplification. Ananalyst might consider possible narratives A → B → Cand D→ E→ F, offering agents a total of two possiblehistories. But if the analyst decides that B could alsolead to F and E to C, the number of possible trajectoriesdoubles, without defining any additional events.

The amount of combinatorial amplification of thenarratives explicitly defined by the analyst depends onthe length of the analyst’s individual narratives and thenumber of interconnections among them, but we can getan idea of the possibilities. Because of the START andSTOP nodes, the CEG is an irregular directed lattice.4

To calibrate our intuitions, consider the paths betweendiagonally opposite corners in a square directed latticeof side n. Such a lattice contains (1+n)2 nodes. Exceptfor the START and STOP nodes, which number only 2for any sized lattice, this is the number of event types

4 In fact, SCAMP does allow causal cycles, so that the CEGno longer defines a partial order over events, but for the pur-pose of this illustrative computation we exclude such cycles.They only add to the space of generated narratives.

10 Parunak et al.

0 2 4 6 8 10 12 14 16

0

50

100

150

200

250

Indegree

Outdegree

Degree

Num

ber

of

Event

Types

Fig. 3 Degree Distribution in Challenge 3 CEG

that the analyst must define. The average node degreein a square lattice asymptotically approaches 4, and bysymmetry in-degree = out-degree = 2.

A simple counting argument [8] shows that the num-ber of simple paths in such a structure is defined bythe central binomial coefficients, 2nCn. A lattice of 441nodes (n = 20) thus generates more than 1.3E11 pos-sible trajectories, each of length 2n = 40. The analystneeds to conceptualize only enough narratives to gen-erate the desired number of event types, with enoughoverlap to link them into a lattice. For a square latticewith indegree = outdegree = 2, each event type needsto appear on average in two narratives. Thus 22 nar-ratives of length 40 (2 ∗ 441/40) covering 441 distinctevent types suffice to yield a 441 node lattice, far fewerthan the 1.3E11 trajectories such a lattice contains.

Our Challenge 3 model has 467 event nodes (in-cluding START and STOP) and an average degree of4.7. The event indegree and outdegree distributions arehighly skewed (Fig. 3), and nearly identical. The eventswith no outgoing edges are STOP nodes for each group,and one event (START) has no incoming edges. Thisskewing will reduce the generative power of the CEG,but even so the number of possible paths greatly ex-ceeds those that the analyst constructing the CEG canexplicitly consider. SCAMP’s swarming ghosts developa probability field over this massive space, sampling itfor data generation (in Ground Truth) or for intelli-gence analysis.

Most events are restricted to agents of one or a fewspecific groups. Events record which groups have agencyfor them. Thus the narrative space is partitioned intosmaller subgraphs for each group. However, agents canmove from one subgraph to another, if they are affili-ated with both groups.

Fig. 4 shows the CEG for the Conflict World. Colorsreflect the agency of the various events. This CEG gen-erates a huge number of alternative narratives, givinga very rich event space within which agents move.

The START node for all groups is at the upper leftof Fig. 4, and has edges to all of the group subgraphs.When an agent reaches the END node, or some othernode that (because of prevent influence edges) offers nonext choices, it returns to START, but in most caseswill not retrace its previous path, since its own stateand the state of the event nodes will have changed. Eachtime the overall participation of an event type nodegoes from 0 to non-zero, a new instance of that eventtype has begun, and ends when the node’s participationdrops again to zero.

Most events unfold over a period of time, and anagent’s participation in an event will generally takesome time before it moves to the next. However, thedelay imposed by an event on different agents will notnecessarily be the same. To model this effect, we assigneach event a nominal transit time describing howlong agents participate in it before moving on. In theabsence of more detailed knowledge, we assume thatevents are Poisson distributed, which means that theirinter-arrival time follows an exponential distribution.So each agent samples its individual transit time froman exponential distribution with the event’s transittime as the parameter. Each time an agent completesan event, it increments its local time by its transittime, and does not execute again until all agents withlower agent times have executed. In our current model,agents do not take the transit time of an event intoaccount in deciding whether to participate in it, a formof the duration neglect bias.

Event participation has other effects on agents inaddition to advancing their clocks.

Fig. 4 Causal Event Graph for Tharum Conflict World

SCAMP’s Stigmergic Model of Social Conflict 11

– If the event involves geospatial movement (Sec-tion 7), the agent’s participation moves it spatially.

– The event’s exogenous features modify the agent’soverall physical, emotional, and economic wellbeing,which in turn modifies the agent’s preferences forthose features, a form of learning. For example, anagent that participates in many events with highpositive economic wellbeing features will tend to re-duce its preference for that feature.

– SCAMP’s social dynamics (Section 8) modulateits preferences based on the other agents whom itmeets, allowing a form of conformity bias.

In addition to agency edges, the CEG also has influ-ence edges. Agents do not move over influence edges. In-fluence edges can connect events in different subgraphs,and modulate the availability of the destination eventand the probability of its selection to agents consider-ing it, based on the degree of participation (the totalpresence features) on the source event. Influence edges,along with HGNs and geospace (Sections 6, 7), allowdifferent groups to interact in SCAMP.

Modelers construct events and edges among themin CMapTools [15], and record event parameters in asheet of the Excel workbook.

5.2 Data from the CEG

We originally planned to disclose full event names to theresearch teams, but DARPA was concerned that theseteams might focus their attention on the semantics ofthe event names, and miss the true causal influences,such as the interaction of features and preferences andthe role of influence edges. In addition, many full eventnames disclose elements of our ground truth, such asthe names of groups, the groups having agency for anevent, or critical locations. So we developed two levelsof reducing the information in the event identifiers, aprocess we call “cheshiring” (inspired by the successivedisappearance of details of the Cheshire Cat in Alicein Wonderland). The level with the least semantics isa simple event number, e.g., E24, for each event. Theintermediate level is a short event name that obscuresthe identity of groups, locations, and key resources. Forexample, “Government security forces arrest minorityleader” becomes “Activity 85 takes place,” “government& armed opposition forces cease negotiations” becomes“negotiations cease,” and “QOL [Quality of Life] at IDP[Internally Displaced Person] camp improves” becomes“QOL at Location 21 improves.” (Our persona allowedus to motivate the imprecision in cheshired event nameson the grounds that the non-English languages spo-ken in Donglap and Tharum were unfamiliar and un-

common, and that the Defense Foreign Language In-stitute was slow in responding to our request for a lin-guistic consultant.) In Challenge 1, we disclosed onlyevent numbers and gave the research teams access toan event’s short name once they had seen informationabout all of its neighbors in the CEG. By Challenge 3,we released the short event names initially, but with acautionary note not to rely on their (very sparse) se-mantic contents.

The main evidence about events available to theresearch teams is an agent history reporting, for eachagent, the event in which it is participating at succes-sive times. For example, Table 2 shows the trajectoryfor an unaffiliated military agent, that is, the history ofevents in which it participates. Each agent maintainsits history internally, but after each agent movement,we also log the move in a file for subsequent analysis.This trajectory is meaningful with full event names, butwith only the short names as requested by DARPA, thetrajectory must be combined with other information todiscern the underlying causality.

Even though different agents in a group can affili-ate differently with other groups, the similarity amongthem can still be detected by comparing the similarityamong their trajectories. One way to do this is withthe “string edit distance,” which is the minimum num-ber of changes (additions, deletions, or replacements ofone element by another) needed to change one string ofcharacters (or list of event types) into the other. Thismeasure, the “Levenshtein distance” [17], can be com-puted in time proportional to the product of the lengthsof the two words by an iterative algorithm due to Wag-ner and Fischer [48].

Fig 5 shows the result of plotting such similaritiesfrom a run in Challenge 1, using multidimensional scal-ing (MDS). We use Kruskal’s nonmetric version [7], asimplemented in R’s isoMDS function.

The plot clearly reflects the impact of the per-groupparameters that define how much each agent’s prefer-ences vary from the group baseline, and how tolerantagents in the group are in affiliating with other groups.Table 1 shows the variation and affiliation thresholdfor each group. Groups with low sampling variation andhigh thresholds generate very similar trajectories, whilethose with higher variation and lower threshold gener-ate more diverse trajectories.

6 Goals

The exogenous and presence features on event nodessupport tactical decisions by agents, based on their pref-erences for these features. Hierarchical goal networks

12 Parunak et al.

Table 2 Example event trajectory for unaffiliated military agent

Domain Time Event Number Event Name Short Name

1 E521 military perceives threat to peace and se-curity threat to peace and security is perceived

17 E536 military implements operational plans operational plans are implemented

28 E351 government sets up checkpoints at officialborder crossings Resource 12 set up at Location 50

143 E24 govt & armed opposition forces increasefighting in border regions Activity 112 increases in Location 50

209 E281government & armed opposition forceswage fierce battle for control of critical ter-ritory

control of critical Resource 110 under con-tention

409 E281 Agent dies Agent dies

(HGNs) for each group modulate the urgency features,supporting strategic decisions.

6.1 Computing over HGNs

An HGN is a directed acyclic graph (not necessarily atree) of goals with labeled edges (and and or) with asingle root goal. Fig. 6 shows the top two levels for theGovernment HGN; the entire HGN has 27 goals with amaximum depth of four goals. A goal’s satisfaction iscomputed from the satisfaction of its subgoals (those atthe origin of edges terminating on it). If the subgoalsconnect through or edges, the satisfaction of the higher-level goal is the maximum of the satisfaction of thesubgoals, while an and delivers the minimum. In thisfragment, the satisfaction of the government’s top-levelgoal 2095 “maintain continuity of rule” is the minimum

−20 −10 0 10 20 30

−30

−20

−10

010

20

30

Gov

AO

RA

Peo

VE

Neu

Fig. 5 Trajectory Similarities

Fig. 6 Top of Government HGN

of the satisfaction of goal 2090 and the maximum ofgoals 2087 and 2088. The urgency of the root is 1 –satisfaction, and propagates down to subgoals.

Every HGN has leaf subgoals, those with no furthersubgoals. These derive their satisfaction from the to-tal presence features on events in the CEG to whichthey are zipped. They can be zipped to any events, notjust those for which the HGN’s group has agency. ThusHGNs provide a second mechanism (in addition to in-fluence edges) for interaction among groups. Zippingcan either support (add to) or block (diminish) the sat-isfaction of the leaf goal. For instance, a leaf subgoalunder 2087 “maintain territorial control” is 2034 “main-tain military superiority.” (Goal 2034 is at the bottomof the government HGN, and so does not appear inFig. 6). Goal 2034 has six zips from the CEG. Threeevent types support it, two from the Government sub-graph and one from the Environment:

1. govt defeats armed opposition forces on multiplefronts

2. government defeats violent extremists on multiplefronts

3. foreign governments provide military support togovernment

Three event types block it, two from the military andone from the armed opposition:

1. significant portion of the military deserts their posts

SCAMP’s Stigmergic Model of Social Conflict 13

Fig. 7 Satisfaction levels in an early experiment

2. military strength weakens3. armed opposition forces kill large numbers of govt

forces

The HGNs thus update an event type’s urgency featuresfrom its presence features.

Though HGNs are specific to individual groups,they can impact agents in other groups. Each group’spreferences can include not only its own urgency, butalso the urgency of other groups (so that it can actto support or block them), and the actual preferencethat an agent shows for a given group’s urgency isthe weighted average of all groups with which it isaffiliated.

Like the CEG, an HGN for each group is constructedby analysts in CMapTools. A tab in the Excel spread-sheet specifies which events are zipped to which goals,and whether they support or block those goals.

6.2 Group Satisfaction over Time

The research teams received goal information in theform of satisfaction levels for each individual agent,which is the average of the root satisfactions of thegroups with which the agent is affiliated, weighted byaffiliation strength. Some agents in each group had noaffiliations, and so presented pure group satisfactions.

Fig. 7 shows an example of how satisfaction lev-els varied over time for different groups in an earlyexperiment (in a version of the model without a dis-tinct Military group). The Armed Opposition rapidlygains satisfaction, but then becomes more frustrated asGovernment satisfaction increases. Relief Agencies takelonger to achieve their goals, but as long as they do, thePeople achieve some satisfaction. However, as the sat-isfaction of Armed Opposition decreases, so does thatof Relief Agencies, and then that of People.

7 Geospatial Constraints

An important dimension of causality is the spatialmovement of agents, initiated by some events in theCEG.

7.1 Geospatial Events and Movement

Fig. 8 shows a map of Tharum and the surroundingcountries. The different colors represent meaningful re-gions, including cities, borders, water features, roads,and (small circles) camps for displaced persons both in-side and outside Tharum, and are defined in the model’sExcel spreadsheet.

A hexagonal grid on the map defines tiles betweenwhich agents move. Our map is 400x400 km, compara-ble to Syria, and each tile is 20km across.

Each agent is initialized at a geospatial location de-pendent on its home group. For example, Violent Ex-tremists start in Muqaa and unofficial border areas (themagenta region just east of Muqaa), while Governmentagents start in large cities (Sag X ).

Some events have specific geospatial destinations.For example, the destination of event type “people goto IDP camp” is a region identifying camps for inter-nally displaced persons (IDPs). We call event types withdestinations, “geospatial events.” When an agent partic-ipates in such an event, it must move through geospacefrom its current location to the event’s destination inorder to complete the event. In this case the durationof its participation on the event is determined, not bythe event’s transit time, but by the time it takes theagent to reach the destination.

Movement through geospace uses the same feature-preference mechanism used in event space, but with areinterpretation of the three kinds of features.

1. The exogenous features reflect difficulty of move-ment, from a terrain map.

2. SCAMP constructs a gradient field over the entiremap for each region or location that can serve asa destination, and when an agent participates in ageospatial event, the gradient for that event’s desti-nation serves as the urgency features that it follows.

3. As a ghost moves from one tile to another, it de-posits presence features for its group not only inthe tile it currently occupies, but also in surround-ing tiles, allowing other ghosts to respond to itspresence. Thus agents interact stigmergically in geo-space just as they do in event space.

The route the agent takes and the time it requiresthus depend on the terrain, its destination, and its at-

14 Parunak et al.

Fig. 8 Map of Tharum and neighbors Quog, Muqaa, and Tlancy

traction to or repulsion from other agents who are alsoin geospace.

Modelers prepare the geospatial map in GIMP [22],using separate layers for distinctive regions, and save itin the OpenRaster (ora) format. The layers are identi-fied in the Excel spreadsheet for the model.

7.2 Examples of Geospatial Movement

Consider an example of agent movement under theseconstraints. A10263 is a Neutral agent affiliated withRelief Agencies, and participates in E211, “relief agen-cies send teams to affected areas to determine imme-diate needs.” The destination for this geospatial eventconsists of large cities, which include Sag Ptulqum, SagPtorg, and Sag Julip. A10263 has a negative preference(-0.29) for Armed Opposition agents (blue), but a pos-itive preference (0.99) for Government agents (orange).Fig. 9 shows the trajectory of A10263 (purple) as itexecutes this mission. Each purple dot represents theagent’s location on a successive day. Its starting loca-tion on the west of Sag Toc is closer to Sag Ptulqum

and Sag Ptorg than to Sag Julip, so it moves north-ward, along a trajectory that commits to neither (left).It senses and moves away from the presence features leftby Armed Opposition, keeping it from approaching SagPtorg. That repulsion, and attraction for Governmentagents (center), leads it to the west, and then north toreach Sag Ptulqum (right).

In addition to this overall trajectory, note how nar-row the path of A10263 is. This agent is affiliated withRelief Agencies, and its preference for this group’s pres-ence features is positive (0.29), so its ghosts, who areplanning its path, are attracted to each other, and tendto form a more focused plan for it to follow.

SCAMP allows the modeler to tune the impactof presence features on ghost movement in geospace.(Avatars still follow their ghosts’ pheromone field.)Fig. 10 shows another run of the configuration inFig. 9, with ghost presence preferences set to 0. NowA10263 ignores the presence of both Government andArmed Opposition, and its ghosts have no inclinationto bunch together, but instead explore more widely.This difference has two implications.

SCAMP’s Stigmergic Model of Social Conflict 15

Fig. 9 Agent (Purple) evading Armed Opposition (Blue) and attracted to Government (Orange) while moving toward SagPtulqum or Sag Ptorg

Fig. 10 Agent with no guidance from presence features

1. They discover a hex where the gradient for Sag Julipis stronger than the gradients for the other two largecities turned off, so they guide the agent to the east.

2. Because their guidance is more diffuse, the avatar’spath is not direct, but more erratic.

8 Social Dynamics

SCAMP agents interact socially, and can change theirbasic group membership based on these interactions.

8.1 Social Interactions and Group Changes

Agents in SCAMP interact socially in three ways.

1. Though each agent has a home group, it can affili-ate with other groups, based on similarity of prefer-ences.

2. An agent can encounter other agents in event space(that is, the CEG) by participating in the sameevent type concurrently, or in geospace by being onthe same tile concurrently. It builds relationshipswith other agents that it meets, relationships that

increase in strength with the number and durationof encounters. These relationships modulate its ownpreferences based on the preferences of those withwhom it interacts [4], implementing a form of group-think or conformity bias.

3. It constantly compares its actual probability of in-teracting with other agents with the probability ofinteraction it would expect based on group affilia-tions. If it finds itself behaving more like members ofanother group, it adjusts their urgency preferencesto favor the goals of that group, and seeks to changeits home group.

In keeping with the stigmergic model, agents do notexchange messages remotely with one another. But theydo have access to the preference vectors of agents withwhom they are colocated on the CEG or in geospace.

The group change mechanism can be attached eitherto event types or to geospatial locations. When an agentbegins participating in such an event type or enters sucha location, it triggers the change process, which may af-fect it or other agents, either collocated with it or else-where. For example, the participation of Governmentagents in Event 113 “government imposes its ideology orreligion on its people” probabilistically triggers a groupchange that changes People who are participating inEvent 492 “people radicalize” into Violent Extremistsparticipating in Event 442 “violent extremists increasetheir numbers,” and People who are participating inEvent 18 “people arm themselves” into members of theArmed Opposition participating in Event 295 “armedopposition forces increase their numbers.” Since Events442 and 295 have high urgency for the Violent Extrem-ists and Armed Opposition, respectively, People agentsthat develop a preference for one of those groups will

16 Parunak et al.

be attracted to these events, increasing the likelihoodthat they will change groups.

The group change mechanism also supports the vir-tual group Guf (in Jewish mysticism, the repository ofsouls). Executing a group change rule from Guf to agroup adds a new agent to that group (for example,when a group recruits a foreigner, not previously rep-resented in the model), while a change from a group toGuf removes an agent from the simulation (for example,a combat fatality). In a violation of the Jewish tradi-tion, avatars do not exist in SCAMP’s Guf before birthand after death, but are instantiated and deallocatedas needed. Guf is simply the name used in in a groupchange rule in place of a group to indicate a birth ordeath event.

Group change rules are encoded in the model’s Excelworkbook.

8.2 Examples of Population Change

The result of the group change mechanism is visible inchanging populations of the different groups over thecourse of a run. Fig. 11 shows the population levels inour Challenge 3 data. It shows several interesting andrealistic features.

– The model defines death events (group change toGuf) for most groups, but birth events (changesfrom Guf) only for Violent Extremists and ArmedOpposition. So the other groups lose population overtime

– VE and AO do grow, but not uniformly. AO startsgrowing first, but then plateaus. Because VE is notpresent at the start of the war and originates outsidethe country, it does not begin to attract recruits atthe start of the model.

Fig. 11 Group populations over time

– Once VE begins to grow, AO also begins to grow.One can think of increasing competition for Peopleinterested in becoming more active in the conflict.

– For a while, VE and AO grow at about the samerate, but VE eventually outpaces AO and ends upwith the stronger force.

9 Challenge Tests

Each challenge involved three tests: Explain, Predict,and Prescribe [20].

9.1 Explain

The Explain test was based on a causal graph, ne-gotiated with the Test and Evaluation (T&E) team(Fig. 12). The nodes for Challenges 1 and 2 are white,while those for the geospace (yellow) and social (pink)perspectives were added for Challenge 3. These nodesare completely independent of the nodes in any ofSCAMP’s perspectives. Per T&E, we divided the nodesinto three categories: the individual Actor, Groups ofactors, and the System within which actors operated.Our ground truth does not include any details of thepolyagent distinction between ghosts and avatars,since that is a computational mechanism that does notcorrespond to any natural social phenomenon or anydata that we gave the research teams.

In Challenges 1 and 2 (the white nodes), the basiccausal loop starts from node 8 (choose next event) to 7(current event) (that is, choosing an event changes thecurrent event). The loop continues to node 10 (eventadjacency in the CEG), an important component ofevent eligibility (node 9a), which returns to 8. Otherloops modulate both agent preferences (to the right)and features of events (to the left). The agency edgesin the CEG implement the edge from 7 to 10, determin-ing from the current event which events are accessiblenext.

In Challenge 3, agents participating in a geospatialevent (node 7) drop into geospace (node 16) at theircurrent location (node 18), making adjacent locations(node 22) accessible (node 9b) and leading to node 23(choose next location), which in turn updates currentlocation (18). When the agent reaches its destination(node 17a) or expends a limiting number of steps (node24), it returns to event space to complete the currentevent (node 7). The pink nodes, representing the socialperspective, modulate agent preferences based on otheragents with whom they interact and enable agents tochange their home group.

SCAMP’s Stigmergic Model of Social Conflict 17

28 Realized

Network

23 Choose

Next Location

17a

Geomvmt

Succeeds

39 Agent’s

Change Group

Preference

33 Within-

Group

Network

36 Agent’s

Actual

P(interact)

31 Generalized

Other of Each

Group

40 Cause

Agent

Transition

22 Location

Adjacency

26 Dest

Preferences

37 Event’s

Change Group

Feature

27 Implicit

Network

32 Group

Population

16 Start

Geomvmt

25 Terrain

Preferences

1 Goals

Group

System

Actor

21 Location

Presence

Features

20a Location

Destination

Features

20b Location

Terrain

Features

9b Location

Eligibility

10 Event

Adjacency

11 Impact on

Wellbeing

12 Actor

Agency

13 Group

Preferences

14 Wellbeing

Preferences

15a Urgency

Preferences

15b Presence

Preferences

17b

Geomvmt

Fails

18 Current

Location

19 Destination

2 Event

Urgency

Features

24 # Geo

Steps

3 Event

Presence

Features

4 Event Static

Features

5a Soft Influence

5b Hard

Influence

6 Group

Affiliation

7 Current

Event

8 Choose

Next Event

9a Event

Eligibility

29 Between-

Group Network 30 Group Social

Influence Process

35 Agent’s

Perceived

P(interact)

38 Agent’s

Social Influence

Process

41 Individual

Wellbeing

42 Group

Wellbeing

Fig. 12 SCAMP’s Ground Truth

Research teams were scored based on the nodes andedges they identified from this diagram. Both teamsidentified numerous nodes that did not correspond toanything in our ground truth. Table 3 shows the num-ber of nodes in each challenge, and the number of nodesthat we could align, either exactly or (in most cases)approximately, with Fig. 12. Because most node identi-fications were only approximate, reporting the numberof edges successfully recovered is very subjective.

9.2 Predict

As the Fourth MEB, we motivated Predict tests as in-formation to guide upcoming decisions. In the Predicttests, we submitted the behavior of our system to T&Efor comparison with the predictions from the researchteams, but did not ourselves participate in the evalua-tion.

Table 3 Explain Results

Challenge # Nodes Chicago JHU

1 17 1 72 17 1 83 46 9 19

9.2.1 Challenge 1

Here is how we introduced Challenge 1’s Predict test.

Our CO has learned of the analysis you areperforming with the data you have requestedfrom the Donglap conflict, and has asked for yourhelp in some upcoming decisions. The situationhere is volatile and may change in several ways.In some cases, he is considering proactive opera-tions. In others, we are aware of things that maychange in theater, and we must respond appro-priately. In each situation, our operational deci-sions will be affected by our estimate of the likelyimpact of each change.

The data that is most relevant to our deci-sions concerns levels of participation in a rangeof event types that we have identified, and we an-ticipate some of these will vary as a result of thechanges in social, political, or military conditionsthat we may either initiate ourselves, or experi-ence due to factors beyond our control. Our COhas asked us to report, for each possible change:

– For which event types will participation in-crease enough to merit our attention?

– For which event types will participation de-crease enough to merit our attention?

18 Parunak et al.

– Which event types will have essentially thesame level of participation that we are seeingnow?

We defined a sufficient increase or decrease as thechange in participation, divided by the square rootof the original participation plus 1 (to avoid divide-by-zero problems). The denominator is roughly thestandard deviation of a normal distribution around theoriginal level, so the quotient approximates a standardscore. We defined a sufficient increase or decrease asone for which this score was greater than 2 or lessthan -2, while we considered the level unchanged ifit fell between -0.4 and 0.4. We also penalized wrongpredictions, to discourage guessing. We defined thechanges for which we sought predictions:

– Intel has identified event type E7930 as onethat leads to a wide range of undesirableoutcomes, and operations says that we couldcompletely suppress it with appropriatedeployment of our military and diplomaticresources. But if we do suppress it, it iscritical to know how participation in otherevent types in Donglap might change ifactors can no longer participate in E7930.

– We have recently observed that actorssimilar to A9979 are no longer participatingin the society. We do not know why thisis so, and do not expect you to be ableto tell us. They have simply “disappeared”(unfortunately a not-uncommon event inthis part of the world). We need to knowhow this is likely to change the participationlevels of various event types.

– We have also observed a decrease (as muchas 50%, though exact figures are uncertain)in the number of actors strongly similar toA4027. Again, we do not know why this isso, and do not expect you to be able to tellus, but we need to know the impact of thischange on participation levels.

– As a result of rising tensions, actors avoidevents likely to be visited by actors differ-ent than themselves. They apparently per-ceive that it’s just not safe to be around peo-ple who are different. How will this increasedcliquishness impact event participation?

9.2.2 Challenge 2

Though the underlying ground truth in Challenge 2 wasthe same as in Challenge 1, we altered the request forthe Predict test:

Once again, we would be grateful for yourhelp in predicting the likely future of events herein Donglap. We are interested in the levels ofparticipation in certain types of events, and inthe degree of satisfaction that different kinds ofactors are likely to feel. In most of the querieswe pose below, we are interested in the course ofevents from this point on in Donglap itself. Thatis, you have seen data from Donglap through day730. We are interested in how that very samesociety will continue to unfold through day 930.Actor identifiers refer to the very same actorsthat they did for the first 730 days.

We are interested in levels of event partici-pation for the following types of events: E8040,E6146, E1939. For these predictions, the level ofparticipation in an event type at a given timeis the number of actors who are participating inthe event type at that time.

Satisfaction (generated by the HGNs, Section 6) wasgenerated in Challenge 1, but not reported to the re-search teams, who were having difficulty recognizing theimportance of groups goals in the ground truth. So inChallenge 2 we decided to report, unsolicited, a timeseries of this variable at the actor level (where it is theweighted average of the satisfaction levels for the groupswith which the agent is affiliated).

We suspect that our informants are nothomogeneous, and that the level of satisfactionthey feel over time may not be the same acrossactors. We are interested in how their satisfac-tion changes over time. In particular, we wouldlike predictions of actor satisfaction for theindividual actor A3317, for actors like each ofA3711, A7347, and A5939, and for the averagesatisfaction across the entire population.

All of the predictions requested so far varyover time. In addition, we would be interestedin how the satisfaction level of actors similar toA7347 would change if there were twice as manyactors of this sort present in the world, all otherpopulations remaining the same. This predictionshould be based on an alternate world that islike Donglap in all respect except for the pop-ulation change. We seek a single number, theaggregate satisfaction of this class of actors forthe first 500 days of such a world. By aggregatesatisfaction, we mean that if you were to drawa curve of the satisfaction for a typical actor ofthis type over the 500 days, we are looking forthe area under the curve. (A smart-aleck 2nd Ltin S6 mumbled something about “integral” and

SCAMP’s Stigmergic Model of Social Conflict 19

“trapezoidal rule,” but the rest of us boxed himpretty roundly around the ears for showing off.)

An important feature of prediction in complex systemsis the existence of a prediction horizon beyond whichany prediction is no better than random [28]. We sus-pected that the lack of accuracy reported by T&E forthe predictions in Challenge 1 might be due to thiseffect, so we asked the research teams for time seriesrather than point predictions, allowing T&E to bothaccuracy and the window over which that accuracy per-sists.

The last requested prediction (the aggregatesatisfaction) does not vary through time, butplease provide not only your best point pre-diction, but also pairs of predictions defining50%, 90%, and 95% confidence intervals. Forthe other predictions, which vary through time,please provide predictions as a daily time series.Again, we would be grateful for your best pointprediction and pairs of predictions defining50%, 90%, and 95% confidence intervals.

9.2.3 Challenge 3

We posed three predictive questions in Challenge 3.Again, we asked for time series rather than point predic-tions, permitting T&E to measure prediction horizon aswell as accuracy. We also included a prediction requestinvolving manipulation of geospace.

Your advice was so helpful in our last de-ployment to Dongglap that we would like to askyour help with some predictions about our cur-rent situation in Tharum.

We would be grateful for three sets of pre-dictions. All three concern the various groups ofagents that you may have discerned are active inTharum.

First, you have seen data from the first 1886days of our deployment. We would like to knowwhat’s coming in the next 100 days. Please giveus time series for this extended period on thetotal population and satisfaction for each of thegroups you have discovered. You may identifythe groups to us by giving us the identifiers forone or more agents whom you believe exemplifythe group, but we want the population and sat-isfaction for the overall group, not for any singleagent you give us.

Second, we expect that E297 is causing trou-ble. In a world that is like Tharum, what woulda time series of the population and satisfactionfor the various groups look like for the first 200

days if we were able to suppress this event, com-pared to what we saw in the first 200 days inTharum?

Third, four regions (shown in color in the at-tached map) also appear to be causing difficulty.Again for a world that is causally like Tharum,what would the population and satisfaction timeseries look like for the first 200 days in a worldwhere we make these regions off-limits to all ac-tors except those who happen to be there at thebeginning?

9.3 Prescribe

In the Prescribe test, the research teams were givendesired outcomes, and asked to prescribe actions thatwould achieve them. Available actions were those thatwe had hypothesized in the Predict tests, for example,

– Exclude an event type from the CEG;– Make a particular spatial region inaccessible;– Change the population of a particular group;– Reduce interaction between specified groups.

The basic methodology in our Prescribe evaluationwas to compare the distribution of five runs of the base-line configuration with five runs of each prescription,and assign two scores: an easy one (how far the me-dian of the prescription runs dominates the median ofthe baseline runs) and a more stringent one (how farthe worst result from the prescription dominates thebest from the baseline runs). We also gave T&E theresults of the best prescription we could devise for eachproblem, based on our full knowledge of the groundtruth. We asked the research teams to achieve the bestimprovement they could with the fewest interventionsand computed gain per intervention, but here we re-port only number of prescriptions that dominate thebaseline, according to the two metrics. We consider aresult to dominate the baseline only if the improvementdivided by the baseline median is strictly positive.

Because SCAMP runs can be time-consuming and(at the time of these tests) we did not have the abilityto cache a run and resume it later, Prescribe runs wereall done on a world with the same ground truth as theinitial data that the research teams received, but start-ing from the beginning rather than from the latest datethat they had seen.

9.3.1 Challenge 1

We tasked the research teams with two problems. Thefirst seeks to minimize or maximize participation onspecified event types:

20 Parunak et al.

We would like to minimize the level of partici-pation of the population in certain destabilizingtypes of events, while increasing their participa-tion in others that we consider favorable. Be-low we list the types of events in each category.Within each list, the event types are listed fromwhat we expect will be the easiest for you to im-pact, to the most challenging. The order belowis our estimate of difficulty. However, we real-ize that your expertise may identify a differentordering.

We asked them to minimize ten events, and maxi-mize 15.

The second problem concerns interactions (concur-rent event participation) between actors from differentgroups:

We observe that the interactions between dif-ferent types of individuals are sometimes helpfulto our mission, while others are detrimental, Wewould like to vary the interaction between ac-tors of different types, based on their similarityto certain designated individuals whom we havebeen tracking. In some cases, we are interestedin alternatives that could either increase or de-crease interaction for the same two types of ac-tors.

Here are the types of actors of interest:– Type A shows behavior like that exhibited

by A1686 around time 3100.– Type B shows behavior like that exhibited by

A4927 around time 2200.– Type C shows behavior like that exhibited

by A9773 around time 2000.Please make separate recommendations to

– increase interactions between actors of typesA and B,

– decrease interactions between actors of typesA and C,

– increase interactions between actors of typesB and C, and also

– decrease interactions between actors of typesB and C.

Table 4 shows the results. A common interventionavailable to the research teams was to exclude certainevents from the CEG. If too many events are excluded,the CEG may become disconnected and the model mayfail to run, so in some cases both teams do not report onthe same number of problems. We report results as frac-tions in which the numerator is the number of prescrip-tions that dominate the baseline while the denominatoris the total number of runnable prescriptions provided.

Table 4 Challenge 1 Prescriptions that Dominate Baseline

Problem ChicagoEasy

ChicagoHard

JHUEasy

JHUHard

Min/MaxPartici-pation

7/22 0/22 24/25 16/25

Inc/DecInteraction 3/4 0/4 4/4 2/4

9.3.2 Challenge 2

The Prescribe test for Challenge 2 built on conceptsfrom the prescriptions requested in Challenge 1 to min-imize or maximize event participation, as well as onthe notion of aggregate satisfaction introduced in theChallenge 2 Predict test. We posed eight problems:

– Maximize participation in E5451– Minimize participation in E8368– Maximize participation in E6739– Maximize participation in E954– Maximize the aggregate satisfaction of agents like

A5205– Maximize the aggregate satisfaction of agents like

A3711– Minimize the aggregate satisfaction of agents like

A7347– Maximize the aggregate satisfaction of agents like

A2180

Table 5 shows the results. There are four problems ofeach type, and all prescriptions were runnable.

9.3.3 Challenge 3

Here is the tasking memo that the research teams re-ceived for the final Prescribe test:

Our CO is considering several alternativetactical objectives in Tharum. We are hopefulthat your knowledge of what is going on in the

Table 5 Challenge 2 Prescriptions that Dominate Baseline

Problem ChicagoEasy

ChicagoHard

JHUEasy

JHUHard

Min/MaxPartici-pation

2 0 1 1

Min/MaxSatisfac-tion

2 0 1 0

SCAMP’s Stigmergic Model of Social Conflict 21

society here would enable you to recommendcourses of action to achieve each of threeoutcomes.

These are independent recommendations: wemight pursue 1, or 2, or 3, and want to know thebest course of action in each case. We do not atthis time seek a course of action to achieve allthree concurrently.

By the time the Prescribe test for Challenge 3 tookplace, COVID-19 was in full swing, so we wove thatinto our request to the research teams, to motivate ourrequest to achieve maximum impact with the fewestinterventions:

As before, the value of a recommendation tous depends not only on the magnitude of theimprovement over the status quo, but also thenumber of interventions required to achieve it.The larger the improvement and the fewer inter-ventions required, the better. (You can think interms of maximizing impact per intervention.) Inaddition, we just received a copy of the Stars andStripes from early March (mail here arrives bycamel train and is quite slow). It reports policiesbeing implemented back stateside that have theeffect of shutting down the society. Such recom-mendations are not practical in Tharum. (Rec-ommendations that try to block too many eventsor locations can have this effect.)

In all cases, we are concerned with groups ofagents of various types. As previously, we defineeach type by identifying a single agent represen-tative of the type, but the questions we wish toaddress in all cases concern groups of agents, notjust these specific individuals

The first of the three problems requires the researchteams to maximize the population of the Armed Oppo-sition over the Government. The second seeks to max-imize the aggregate satisfaction of Armed Oppositionover the Government, while the third seeks to mini-mize visits by military groups (Government, Armed Op-position, Violent Extremists, and Military) to refugeecamps in other countries.

1. The US Government wishes to supportagents like A6313, particularly in compar-ison with agents like A1854. What stepsshould we take so that by day 130, thedifference in total population of the firsttype over that of the second type is maxi-mal? Recommendations of the form “Reducethe number of agents similar to A1854” or“Increase the number of agents similar to

A6313” are too vague to be useful to us inthis case.

2. Recall the concept of aggregate satisfactionfrom our prescription request for Donglap.We wish to maximize the aggregate of thedifference in satisfaction between agents likeA6313 and those like A1854. How can we dothis?

3. We have learned that the following locationsin neighboring countries are crucial tohumanitarian relief: Locations 10x3, 11x2,14x4, 15x3, 20x20, 21x19, and 21x20. Butthe presence of military forces in these areas(agents like A6313, A1854, A3779, andA2763) is hindering this activity. What canwe do to reduce the total number of visits ofagents of these four types to these locationsover the first 130 days? NB: One actionavailable to us within Tharum is to blockadea region of interest to prevent agents fromentering it, but since we do not have a Statusof Forces agreement with the neighboringcountries, this type of operation is notavailable to us in pursuing this objective.

The third problem illustrates how our military per-sona provides a rationale for asking for prescriptionsthat requires an understanding of the relation betweenevents and geospatial locations, rather than allowingthem simply to exclude the locations listed.

Table 6 shows the results. All prescriptions wererunnable on all three problems.

10 Summary

Conflict World exhibits several important features ofSCAMP.

– The entire model is constructed by domain experts,professional analysts who are not programmers, us-

Table 6 Challenge 3 Prescriptions that Dominate Baseline

Problem ChicagoEasy

ChicagoHard

JHUEasy

JHUHard

Max PopAO/Gov 0 0 0 0

Max SatAO/Gov 0 0 1 0

MinVisitstoCamps

1 0 0 0

22 Parunak et al.

ing three common desktop tools: the CMap conceptmapper, GIMP, and Excel. Thus it can easily cap-ture their insights without the knowledge acquisi-tion barrier between domain experts and program-mers common in many applications, and allows usto build models of realistic complexity.

– The resulting model artifacts give an unambiguouscausal ground truth that underlies the data given tothe research teams, and is easily modified to exam-ine the causal impacts.

– In spite of stigmergy’s simplicity, Simon’s Lawenables SCAMP to capture psychologically andsocially realistic dynamics [25,32], including de-liberate tactical choice guided by preferences overalternatives, non-deterministic decision-making [3],strategic (goal-driven) as well as tactical decisions[40], use of mental simulation to look ahead in time[16], interactions with other agents encounteredon events or geospatial tiles as a mechanism foradjusting individual preferences [11], the centralityof narrative as a mental representation [10], andnaturally bounded rationality [42].

Our experience in the program led to several lessons.

– Adopting a detailed backstory and persona was veryhelpful in managing our interactions with the re-search teams.

– The requirement repeatedly urged by T&E not togive away our ground truth often made it difficultto respond to questions that in the real worldwould have been legitimate. It would be interestingto consider alternative ways to validate causalitythat avoid this constraint. For example, we mightgive the research teams our ground truth togetherwith a number of data sets, some generated bythe ground truth and some by increasingly severeablations of it, and see how well they could identifythe real ones.

– Different disciplines in the social sciences talk aboutthe world in different ways. The lack of a shared on-tology for psychological and sociological phenomenawas a major hindrance in our communication withthe research teams.

Interested researchers can obtain the model underthe Gnu Public License, and a detail user manual, de-scribing the construction of the various configurationfiles, is available [33].

Acknowledgements This research was funded by the De-fense Advanced Research Projects Agency (DARPA), underCooperative Agreement HR00111820003. The content of thispaper does not necessarily reflect the position or the policyof the US Government, and no official endorsement should beinferred.

References

1. Argonne National Laboratory: Repast agent simulationtoolkit (2007). URL http://repast.sourceforge.net/

2. Bonabeau, E., Dorigo, M., Theraulaz, G.: Swarm Intelli-gence: From Natural to Artificial Systems. SFI Studiesin the Sciences of Complexity. Oxford University Press,New York (1999)

3. Busemeyer, J.R., Townsend, J.T.: Decision field theory:A dynamic-cognitive approach to decision making in anuncertain environment. Psychological Review 100(3),432–459 (1993)

4. Camazine, S., Deneubourg, J.L., Franks, N.R., Sneyd, j.,Theraulaz, G., Bonabeau, E.: Self-Organization in Bio-logical Systems. Princeton Studies in Complexity. Prince-ton University Press, Princeton, NJ (2001)

5. Álvaro Carrera, Iglesias, C.A.: B2di - a bayesian bdiagent model with causal belief updating based on msbn.In: Proceedings of the 4th International Conference onAgents and Artificial Intelligence - Volume 2: ICAART„pp. 343–346. INSTICC, SciTePress (2012). DOI 10.5220/0003746003430346

6. Cioffi-Revilla, C.: Introduction to Computational SocialScience, 2nd edn. Texts in Computer Science. Springer,Cham, CH (2017)

7. Cox, T.F., Cox, M.A.A.: Multidimensional Scaling, sec-ond edn. Chapman & Hall (2000)

8. Dickau, R.: Counting lattice paths (2020). URL https://www.robertdickau.com/lattices.html

9. Epstein, J.M.: Generative Social Science. Princeton Stud-ies in Complexity. Princeton University Press, Princeton,NJ (2006)

10. Fisher, W.R.: Human Communication as Narration: To-ward a Philosophy of Reason, Value, and Action. Uni-versity of South Carolina Press, Columbia, SC (1989)

11. Friedkin, N.E., Johnsen, E.C.: Social Influence NetworkTheory: A Sociological Examination of Small Group Dy-namics. Structural Analysis in the Social Sciences. Cam-bridge University Press, Cambridge (2011)

12. Gilbert, N., Troitzsch, K.G.: Simulation for the SocialScientist, 2 edn. Open University Press, Buckingham,United Kingdom (2005)

13. Grassé, P.P.: La reconstruction du nid et les coordi-nations inter-individuelles chez bellicositermes natalen-sis et cubitermes sp. la théorie de la stigmergie: Essaid’interprétation du comportement des termites construc-teurs. Insectes Sociaux 6, 41–84 (1959)

14. Grimm, V., Berger, U., DeAngelis, D.L., Polhill, J.G.,Giske, J., Railsback, S.F.: The odd protocol: A re-view and first update. Ecological Modelling 221(23),2760–2768 (2010). URL http://www.agsm.edu.au/bobm/teaching/SimSS/ODD_protocol.pdf

15. IHMC: Ihmc cmaptools - download (2013). URL https://cmap.ihmc.us/cmaptools/

16. Kahneman, D., Tversky, A.: The simulation heuristic. In:D. Kahneman, P. Slovic, A. Tversky (eds.) Judgmentunder Uncertainty: Heuristics and Biases, pp. 201–208.Cambridge University Press, Cambridge, UK (1982)

17. Levenshtein, V.I.: Binary codes capable of correctingdeletions, insertions, and reversals. Soviet Physics Dok-lady 10(8), 707–710 (1966)

18. Luke, S., Cioffi-Revilla, C., Panait, L., Sullivan, K.,Balan, G.: Mason: A multiagent simulation environment.SIMULATION 81 (2005)

19. de Marchi, S.: Computational and Mathematical Model-ing in the Social Sciences. Cambridge University Press,Cambridge, UK (2005)

SCAMP’s Stigmergic Model of Social Conflict 23

20. Naugle, A., Russell, A., Lakkaraju, K., Swiler, L., Verzi,S., Keywords:, V.R.: The ground truth program: Simu-lations as test beds for social science research methods.Computational and Mathematical Organization Theory(forthcoming) (20210)

21. Parunak, H.V.D.: ’go to the ant’: Engineering principlesfrom natural agent systems. Annals of Operations Re-search 75, 69–101 (1997)

22. Parunak, H.V.D.: A survey of environments and mech-anisms for human-human stigmergy. In: D. Weyns,F. Michel, H.V.D. Parunak (eds.) Proceedings of E4MAS2005, Lecture Notes on AI, vol. LNAI 3830, pp. 163–186.Springer (2006)

23. Parunak, H.V.D.: Between agents and mean fields. In:J. Sabater, D. Villatoro, J. Sichman (eds.) the Work-shop on Multi-Agent-Based Simulation (MABS 11) atAAMAS 2011, LNAI, pp. 106–117. Springer, Taipei, Tai-wan (2011)

24. Parunak, H.V.D.: Odd protocol for scamp. Report,Wright State Research Institute (2020). URL https://abcresearch.org/abc/papers/ODD4SCAMP.pdf

25. Parunak, H.V.D.: Psychology from stigmergy. In: Com-putational Social Science (CSS 2020), vol. (forthcoming).CSSSA (2020)

26. Parunak, H.V.D.: How to turn an mas into a graphicalcausal model. Autonomous Agents and Multi-Agent Sys-tems p. (under revision) (2021)

27. Parunak, H.V.D.: Social simulation for non-hackers. In:K.H. Van Dam, N. Verstaevel (eds.) 22nd InternationalWorkshop on Multi-Agent-Based Simulation (MABS2021). Springer (2021)

28. Parunak, H.V.D., Belding, T.C., Brueckner, S.A.: Predic-tion horizons in agent models. In: D. Weyns, S. Brueck-ner, Y. Demazeau (eds.) Engineering Environment-Mediated Multiagent Systems (Satellite Conference atECCS 2007), vol. LNCS 5049, pp. 88–102. Springer(2008)

29. Parunak, H.V.D., Brueckner, S.: Concurrent modeling ofalternative worlds with polyagents. In: the Seventh In-ternational Workshop on Multi-Agent-Based Simulation(MABS06, at AAMAS06), pp. 128–141. Springer (2006)

30. Parunak, H.V.D., Brueckner, S.A.: Engineering swarmingsystems. In: F. Bergenti, M.P. Gleizes, F. Zambonelli(eds.) Methodologies and Software Engineering for AgentSystems, pp. 341–376. Kluwer (2004)

31. Parunak, H.V.D., Brueckner, S.A., Matthews, R., Sauter,J.: Swarming methods for geospatial reasoning. Interna-tional Journal of Geographical Information Science 20(9(Oct)), 945–964 (2006)

32. Parunak, H.V.D., Morell, J.A., Sappelsa, L.: The scamparchitecture for social simulation (2020). In preparationfor JASSS

33. Parunak, H.V.D., Morell, J.A., Sappelsa, L., Greanya,J.: Scamp user manual. Report, Parallax Advanced Re-search (2020). URL https://www.abcresearch.org/abc/papers/SCAMPUserManual.zip

34. Parunak, H.V.D., Purcell, M., O’Connell, R.: Digitalpheromones for autonomous coordination of swarminguav’s. In: First AIAA Unmanned Aerospace Vehicles,Systems,Technologies, and Operations Conference. AIAA(2002)

35. Parunak, H.V.D., Savit, R., Riolo, R.L.: Agent-basedmodeling vs. equation-based modeling: A case study andusers’ guide. In: N. Gilbert, R. Conte, J.S. Sichman(eds.) Multi-Agent Systems and Agent-Based Simula-tion, First International Workshop, LNCS, pp. 10–25.Springer, Berlin, Germany (1998)

36. Pynadath, D.V., Dilkina, B., Jeong, D.C., John, R.S.,Marsella, S.C., Merchant, C., Miller, L.C., Read, S.J.:Disaster worlddecision-theoretic agents for simulatingpopulation responses to hurricanes. Computational& Mathematical Organization Theory (forthcoming)(2021)

37. Rager, S., Leung, A., Pinegar, S., Mangels, J., Poole,M.S., Contractor, N.: Groups, governance, and greed:The access world model. Computational & MathematicalOrganization Theory (forthcoming) (2021)

38. Rao, A.S., Georgeff, M.P.: Bdi agents: From theory topractice. In: the First International Conference on Multi-Agent Systems (ICMAS-95), pp. 312–319. AAAI (1995)

39. Sappelsa, L., Parunak, H.V.D., Brueckner, S.: Thegeneric narrative space model as an intelligence analysistool. American Intelligence Journal 31(2), 69–78 (2014)

40. Shivashankar, V.: Hierarchical goal networks: Formalismsand algorithms for planning and acting. Ph.d., ComputerScience (2015)

41. Simon, H.: The Sciences of the Artificial. MIT Press,Cambridge, MA (1969)

42. Simon, H.A.: A behavioral model of rational choice. TheQuarterly Journal of Economics 69(1), 99–118 (1955)

43. Sterman, J.: Business Dynamics. McGraw-Hill, NewYork, NY (2000)

44. Sullivan, K., Coletti, M., Luke, S.: Geomason: Geospatialsupport for mason. Tech. rep., George Mason University(2010)

45. Sumpter, D.J.: Collective Animal Behavior. PrincetonUniversity Press, Princeton, NJ (2010)

46. Team, T.G.: Gimp: Gnu image manipulation program(2020). URL www.gimp.org

47. Vidal, J.M., Durfee, E.H.: Recursive agent modeling us-ing limited rationality. In: First International Conferenceon Multi-Agent Systems (ICMAS 95), pp. 125–132 (1995)

48. Wagner, R.A., Fischer, M.J.: The string-to-string correc-tion problem. Journal of the ACM 21(1), 168–173 (1974)

49. Wilensky, U.: Netlogo (1999). URL http://ccl.northwestern.edu/netlogo

50. Züfle, A., Wenk, C., Pfoser, D., Crooks, A., Kavak, H.,Kim, J.S., Jin, H.: Urban life: A model of people andplaces. Computational & Mathematical OrganizationTheory (forthcoming) (2021)