CONFLICT: Agents in conflict situation(extended abstract of the MSc dissertation)

Henrique Teles CamposDepartamento de Engenharia Informatica

Instituto Superior Tecnico

Advisor: Professor Ana Maria Severino de Almeida Paiva

Abstract—The objective of this document is to addressthe problem of how to create groups of agents that engagein natural conflict situations, inspired by what happens inreal world scenarios. For such, we propose that emotionalprocesses are a key aspect to convey conflict situations’ aspects,such as, emergence and escalation. Further, we intended topopulate an educational conflict resolution game with theseagents. To achieve this, we investigated conflict theory inSocial Psychology. Then we reviewed serious games, where weanalysed their purposes and how they were achieved. Withthat, we proposed a conflict model, where the agent observes asituation, then makes assumptions about it and finally behavestowards it. This model was implemented in a emotionally-driven agents’ architecture. Then, it was integrated in a conflictresolution game demonstrator, the Dream Theatre game. Aworld simulation framework supported the integration. Withthis, we conducted an experiment to assess our model’seffectiveness. The experiment consisted in preliminary testsand a final between groups evaluation. In both, participantshad to watch a recorded video of a user interacting withthe Dream Theatre demonstrator. The final evaluation resultswere consistent with our hypothesis, which states that agentswith a decision-making process which stems from an emotion-oriented architecture will provide a believable display of overtmanifestation of conflicts.

I. INTRODUCTION

Over the past few years, there has been a great increase ofinvestigation in the AI field of Multi-Agents systems (MAS).Particularly, in the field of synthetic characters, researchersare aiming towards greater believability.

On this research, we investigate on how to imbue syntheticcharacters with human-like behaviours towards conflict sit-uations. Therefore, increasing their believability.

Conflict is a natural part of your lives. It is presentin our everyday basis. In our society, conflict has beenconsidered to be something to avoid due to the negativefeelings and destructive behaviours [1]. However, recentresearch acknowledges that conflict can yield beneficialaspects as well [1]. As such, conflict resolution strategiesare essential behavioural skills that all of us need to learnsince early childhood. In fact, many schools now embraceconflict resolution learning schemes to help children toacquire attitudes towards others in solving their divergencesand problems in a positive and constructive manner.

On the other hand, over the years, there has been severalresearch on developing new teaching tools that motivate

and engage today’s children. Information technologies andcomputer and video games have show to be a great tool forteaching [2], educating [3] and raising awareness [4] aboutsubjects on students.

Motivated by the conflict resolution education and theuprising of serious games for teaching children, the SIRENproject [5] intends to create a conflict resolution game.

In this research, we are concerned about the NPCs(agents) that will populate the serious game. In order to dothat, a rather general problem that we try to handle is statedas follows:

How can we create autonomous agents, for aserious game, that are able to identify socialconflict situations with other agents and handlethem in a believable way?

Our research focuses on agents’ behaviours towards con-flict between themselves and how this aspect affects agents’believability. Our approach to solve this problem will beby investigating what are the most important characteristicsof human interactions within a conflict situation. We willaim particularly at what causes conflict, its protagonists, itseffects in individuals and how it can be resolved. Withinthe latter referenced aspect of conflict, we are particularlyconcerned about which strategies individuals use to handleconflict and what factors makes them use such handlingmechanisms and what makes conflicts escalate.

Based on the several aspects of conflict identified in thepreviously described investigation, we will develop a modelfor conflict, which will be integrated in the architecture ofan agents’ society, providing the agents with the ability torecognise conflicts and handle them autonomously. We willalso integrate our agents in a simple serious game scenariothat will be populated with the agents’ society. Thus, withthis research we will try to prove the following hypothesis:

If the agents’ decision-making process stems froman emotion-oriented architecture, their behaviourswill be based on emotional grounds and intensity,which will play an important role in overt mani-festations of conflict and users may recognise suchbehaviours as believable.

Additionally, note that this problem does not focus onagents’ ability to resolve conflicts. This investigation doesnot aim in modelling agents’ strategies to act with and

1

towards other agents in order to achieve a certain goal. Whatwe intend to model is the emotional influence in conflictemergence and escalation, by affecting agents’ behaviours.

Further, as conflict is quite a broad subject, throughoutthis document we will specify which aspects of conflict willbe considered for our research.

II. BACKGROUND - THEORY OF CONFLICT

There is not a reconciled definition for conflict. However,there are some accepted characteristics that allow us tocharacterise the concept in terms of its structure. We may saythat conflict varies along five dimensions. Participants referto individuals that may take part in the conflict situation.Causes are linked to adverse or external conditions thatinitiate a conflict episode, as for example, relationship man-agement, self-interest oriented actions or lack of information.Based on the initiating action, the actor’s responses (one’s at-titudes, behaviours or strategies) will determine the course ofaction. Actors’ behaviours determined to bring the situationto an end are embedded in the conflict resolution dimension,which following Thomas’ taxonomy [6] the approaches aregeneralised as: accommodation, avoidance, competition, col-laboration and compromise. The dimensions of assertivenessand cooperativeness are phrased as intentional terms, inwhich assertiveness refers to the extent to which protagoniststry to achieve their own goal and cooperativeness refers tothe extent of protagonists trying to satisfy the concerns ofothers. Depending on the approaches towards the conflict,outcomes refer to the consequences that conflict bringsabout.

As this description may suggest, the conflict episode is atemporal sequence of events [7], which may be comparedto the typical plot of a narrative. This comparison wasmade clear by Laursen and Pursell [7], who time-distributedthe conflict episode. The conflict-plot is identified by aninitiating action (complication), a rising action (set of actionsthat contribute to conflict escalation), a climax and theoutcome.

When the conflict gets worse, we say that it escalates.When it reaches the turning point (climax) and the mag-nitude of the situation decreases, we say that de-escalates.Every conflict escalates, even if it is only in a unilateralway. Escalation occurs when one or both parties engage inthe conflict, moving it from a less severe stage to a morecontentious and heavy state [8]. According to Pruitt [8], anexample of such situation is persistent annoyance, whichmay lead to a sequential progression of angry statements,threats and eventually the conflict escalation to an extreme,where harassment and abuse occurs. However, what maymake one more prone to escalation and which triggerspromote progression to more aggressive tactics it is notclear in the literature [8]. Nevertheless, we may say thatescalation is driven by inner triggers [9], that is, emotionsthat weight one’s current goals and assess the affective valueof the situation [10].

Due to destructive outcomes in conflict amongst childrenand teenagers, conflict resolution educational has been ap-

plied in school in order to better educate children on conflictawareness and proper handling.

III. RELATED WORK

Our related work centres in games as learning mecha-nisms, serious games. More precisely, we reviewed gamesthat somehow approach conflict situations. Each game wasreviewed as follows (see Table I):

• The Prom – focus on relationship conflicts, wherethe user had to manage social relationships by takingactions that will balance the social world. Conflictsemerged due to characters’ distinct personalities (in-terests, needs, traits, social networks and social status).Further, the user is directly involved in the conflict, ashe may cause and resolve it.

• Global Conflict – concerns conflicts of values (theIsraeli-Palestinian conflict), where players had to framethe perspective of both conflicting parties. To do so,players need to maintain neutrality with both sides.Given the intractable aspect of the conflict portrayedby this game, the player role is merely to gather infor-mation of both sides, as the player does not interferewith the situation.

• FearNot! – handles conflicts of relations in the contextof bullying, where the user takes responsibility for avictim of bullying and has to help him make decisions.Conflicts happened due to the bully’s unreasonableperception and conduct towards the victim. The bully’sactions towards the victim elicit emotions in both ofthem. The player’s role is to advice the victim characteron how to handle such situations, thus, this advices in-fluence how the victim will behave on the next bullyingepisode, hence the user is indirectly involved in theconflict. These advices may improve or deteriorate thevictim’s ability to confront the bully. Therefore, we canassume that whenever the victim learns to effectivelyconfront the bully, the conflict de-escalates. In contrary,deteriorating advices may escalate the conflict.

ConflictGame Type User Involvement Emotional influence Escalation De-escalation

The Prom Relation Yes - Direct No No NoGC: Palestine Values No - Perspective taking No No No

FearNot! Relation Yes - Indirect Yes Yes Yes

Table ICOMPARISON OF CONFLICTS PORTRAYED IN GC: Palestine, The Prom

AND FearNot!.

IV. CONFLICT MODEL

In agents’ societies where agents have incompatible goals,conflicts are bound to emerge. Yet, such conflicts are usuallyundertaken by the agents’ architectures, which try to find away to cooperate with the parties involved. In this work, wetake a different approach by trying to address not so mucha cooperative-based concept but rather give a step towardsa more contingent view of conflict and its resolution.

Emotions are at the heart of social interaction and theyplay a relevant role triggering events such as conflict.

2

Therefore, the cognitive appraisal is an essential elementfor understanding conflict. Given that such situations emergefrom one’s subjective evaluations of the environment.

In this Section we describe a conflict handling modelfor agents, where emotions affect their behaviours towardsthe other participants in the conflict situation. Our modelstems from FAtiMA’s emotional model for agents [11] andin Tessier’s et al. agents’ conflict handling action model [12].

Our model is illustrated in Figure 1 and further describedbelow. The conflict dynamics specified by the model triesto capture the essence of Thomas’ [6] definition of thephenomenon, in which conflict is defined as “the processwhich begins when one party perceives that another hasfrustrated, or is about to frustrate, some concern of his”.Therefore, first the agent has to perceive that some actionperformed by any other agent frustrated (or is going tofrustrate) its goals, then the event is diagnosed not onlyaccording to its congruence to the agent’s goals, but alsoother contextual elements (e.g. social relationships). Finally,behaviours are selected in line with the state of the worldand the agent inward beliefs. To sum up, an agent initiallyobserves, then makes assumptions and finally communicateswith others.

Figure 1. Conflict handling model.

The implemented model works as follows.

A. Conflict Recognition ModuleOthers’ actions or events that affect (positively or nega-

tively) a certain concern of the agent are perceived. Theseare checked by the agent itself to evaluate whether theyraise potential conflicts or contribute to the escalation ofthe current situation. It is specified the urgency of conflictthat determines how intense the situation is.

B. Conflict Diagnosis ModuleThis module diagnosis the conflict oriented event in 3

steps:1) analyses the conflict event and generate a conflict de-

scription – this description consists in what is the cause(goal frustrated), participants involved and relationshipbetween them and the importance of the conflict;

2) generates emotional reactions from the conflict analysis– an emotion reaction is generated in function of whatwas appraised.

3) generates emotions from the emotion reactions, wherethe intensity of the emotion reflects the urgency of theconflict [13], and integrates into the emotional state(this process is undertaken by FAtiMA [11]);

C. Conflict Behaviour Selection Module

As described in Section II, behaviours for handling con-flict range within two dimensions assertiveness and cooper-ativeness and for simplification reasons we only consideredattacking and evading behaviours (from Raider’s AEIOUmodel of communication in conflict [14]), which are asso-ciated to [high assertiveness, low cooperativeness] and [lowassertiveness, low cooperativeness], respectively. The valuesof assertiveness and cooperativeness are balanced by theagent’s emotional state. The reason behind that choice isbased on the assumption that negative emotions are linkedto less cooperative approaches [10], which will lead to moreconflicts and their consequent escalation. For example, anagent becomes less cooperative as he gets more frustratedwith the situation at hands. In this model, whether an agentis more prone to one of the aforementioned behaviours isdetermined by their behavioural predispositions.

D. Behavioural Predispositions

Our current investigation only aims to model simplebehaviours, in order to demonstrate escalation. We are awarethat several kinds of behaviours can be present in destructiveconflict situations [15]. One’s personality, for instance, canhave impact on that matter. However, we aim to explore theevolution of a conflict by the means of overt manifestations.Therefore, we decided to focus on a set of behavioursthat might lead to potential conflicts and their escalationto explore what believable conflicting behaviour might beat the eyes of the human perceiver. In this way, we set thatnegative emotions will affect negatively the agent’s actionstowards conflict diminishing the possibility for cooperation[16]. Section V describes the scenario where this model wasapplied and then tested. A broader and more detailed set ofbehaviours will be developed in future work.

1) Attacking Behaviour: Agents with the tendency toAttack follow a destructive path to cope with the conflict[14]. These agents are prone to have high assertivenessand low cooperativeness. This falls into the competitiveapproach [6] (see conflict handling approaches in SectionII), in which the agent only wants to benefit himself. Theactions taken by agents with this tendency range from a lowlevel of aggressiveness to an extreme. For example, as theagents’ emotional state worsens, their actions may progressas follows: lesser insult, criticise negatively, harsh insult, andthreat.

2) Evading Behaviour: On the other hand, an agent withan Evading tendency may try to avoid conflict situations.Initially an agent with these characteristics may want tocooperate [14], however the build up of negative emotionsas a result of a negative social exchange makes the agentto become less cooperative. Therefore, the agent will havea tendency to apply Avoiding approaches [6] (see conflicthandling approaches in Section II), followed by hostileevasions and withdraw of his interests. This pattern of actioncharacterises the Evading behaviour [14]. Furthermore, asthe emotional state gets worse the actions performed by an

3

agent with this kind of behaviour progresses as follows:ignore the situation, sacrifice own’s goals to avoid furtherinvolvement and, finally, leave the scene.

V. THE DREAM THEATRE GAME

The Dream Theatre1 is an educational game that aimsat teaching children, aged 9 to 11, some conflict resolu-tion skills. The game setting is a theatre company andthe user/child needs to select the adequate cast for eachperformance. Every week the child is challenged with a new“production” and needs to select adequately his/her cast, andassign adequately the roles to the actors. The user has aset of possible virtual actors/characters to assign the roles,and needs to do so, trying to optimise the final performanceresults. The only problem is that the agents may not acceptthe assignment made by the learner, and conflicts may occurbetween the agents. The role of the child is to manage theconflict, advice the agents, and try to do so in a manner thatthe conflict is resolved, otherwise, if the emotional statesof the agents at the end is very negative, the performanceresults will not be good.

Figure 2. Dream Theater’s screenshot, showing two characters having adiscussion about a role.

Agents in the cast have a set of characteristics, in particu-lar they have preferences for roles (for example some agentsmay prefer to be a “Hero”, whereas others may prefer to bea “Villain”). As the player grants roles to characters, conflictsituations emerge when characters perceive an obstruction toachieve their desired roles. They also have social relations(friendship relations, allowing for in-group and out-groupsituations), and proficiency. All the agents are modelledaccording to an agent architecture, where emotions areappraised by the situations created by the user, and theagents reacting emotionally to the situations.

In the current version of our demonstrator, the playerdoes not have specific strategies on approaching the conflictsituation, other than granting roles, dismissing from roles orfiring characters from the theatre company.

1The assets of the game scenario were developed by Serious GamesInteractive - www.seriousgames.dk

A. Agents’ Minds

Our conflict model, described previously, was imple-mented in FAtiMA emotional agents’ architecture [11] (seeFigure 3). No significant changes were made to FAtiMA’sarchitecture, as we took advantage of the reactive appraisalprocess and the deliberative planning process to implementour conflict model in the agents.

Figure 3. Agents’ implementation in FAtiMA’s architecture (diagram from[17]), where the circle notations represent where each conflict model’s stepwas implemented.

Therefore, the model was implemented as follows.1) Conflict recognition and diagnosis: On integrating our

conflict model in FAtiMA agents’ minds architecture, wefound plausible to simplify some of the model’s aspects. Theconflict recognition was integrated into conflict diagnosisdue to the fact that we used FAtiMA’s deliberative layerto implement these steps from our conflict model.

To begin with, the CR is defined by a set of active-pursuitgoals in FAtiMA’s deliberative layer. In FAtiMA, active-pursuit goals are goals that the agents actively try to achieve[11]. The agents plan on achieving such goals whenever theyare activated, i.e., when the goals’ preconditions are satisfied.

In the conflict recognition’s active-pursuit goals definition,the pre-conditions must define the situation in which acertain action is directed to a certain interest of the agent.Furthermore, the goal’s success conditions must specify the“internal action”2 which will cause the conflict diagnosisstep.

The conflict diagnosis step is supported by FAtiMA’sdeliberative planner, which will plan to perform the “internalaction” (FAtiMA’s deliberative planner process’ descriptioncan be found in [17]). For this, it is specified a set of actionsthat determine the analysis of the event, where each actionpresents the conflict description.

With the conflict description, an emotional reaction mustbe generated, according to how the agent agent believes thatthe event will affect his interest. By taking advantage ofFAtiMA’s reactive appraisal [11], this step of the conflictdiagnosis is defined by a set of emotional reactions directedto the conflict descriptions from the “internal actions”.

2We use the term “internal action” to specify an agent’s action which isnot visible to others, only to the one who performs it.

4

After the emotional reaction is triggered, FAtiMA’s ap-praisal process generates the emotions [17], followingOCC’s theory of emotions [18]. This process consists intwo major steps. The first generates potential emotions andthe second will send these emotions to the emotional state.

2) Conflict behaviour selection: In a nutshell, the conflictbehaviour selection step is defined as a set of active-pursuit goals towards the conflict situation, where emotionsinfluence which goals the agent will be capable of activating.

As we stated previously, in our model’s description (Sec-tion IV), we aim at modelling behaviour which is affected bythe agents’ emotional states. With such, for this version ofthe agent’s minds implementation, we use agents’ moods asthe key emotional affect to influence behaviours. However,this assumption can be quite arguable, as we could modelconflict behaviours related to specific emotions, such asanger or fear [10][19]. However, as mood is a pervasiveaffect that defines a general disposition state of an individual,we believe that this emotional aspect is more suitable toinfluence characters’ behaviours towards conflict.

Further, we decided to implement two behavioural pre-dispositions (as presented in Section IV): Attacking andEvading; as we believe these are more likely to generateescalation as a result of what those behaviours bring to thesocial interaction.

The Attacking behaviour has a set of aggressive be-haviours towards the situation of conflict. In Table II, wepresent the Attacking agents’ emotionally-driven behavioursto approach a conflict situation. Each one represents a groupof active-pursuit goals, which can be activated if agents’emotional states are in those mood ranges.

Emotionally-driven behaviour Mood(m) DescriptionDo not mind m ≥ 0 Agent performs a rather “neu-

tral” behaviour towards thesituation.

Insult other −2 < m < 0 Agent uses insults towardsother agents.

Criticize negatively other −4.0 < m ≤ −2.0 Agent uses destructive criticstowards other agents.

Threaten other m ≤ −4.0 Agent threatens the other towithdraw from the threateningagents’ concern.

Table IIAttacking AGENTS’ SET OF BEHAVIOURS, ACCORDING TO MOOD’S

VALUES.

It is also important to explain that this agent’s emotionalreactions are manipulated, so that the agent finds very unde-sirable the others’ attempts to approach the conflict. Hence,the more the conflict situation prolongs, more aggressivethe agent will be, as others’ actions will only deteriorate hisemotional state.

In contrast, the Evading behaviour agent has a set ofless aggressive behaviours towards a situation of conflict.As we referred, in Section IV, Evading agents may initiallyhave constructive approaches towards the conflict situation,however, if the emotional state deteriorates, this agent willbegin to follow evasive approaches. Following that, Table IIIpoints out Evading agents’ emotionally-driven behaviours.

The values for the mood ranges were determined by empir-ical tests with the FAtiMA architecture.

Emotionally-driven behaviour Mood(m) DescriptionDo not mind m ≥ 0 Agent performs a rather “neu-

tral” behaviour towards thesituation.

Ask why −3.0 < m ≤ 0 If approached by anotheragent, this agent will try to askthe reason for such behaviour.

Do not respond −5.0 < m ≤ −3.0 Agent does not fight back ag-gressive moves towards him-self (e.g., deny a critic or ainsult towards himself).

Give up −8.0 < m ≤ −5.0 Agent gives up desired inter-est, in order to avoid any moreinvolvement in the conflict.

Leave scene m ≤ −8.0 Agent final withdrawal, inwhich he leaves the scene.

Table IIIEvading AGENT’S BEHAVIOURS, ACCORDING TO MOOD’S VALUES.

Similarly to the Aggressive agent, this agent’s emotionalreactions were authored for this agent to find undesirableaggressive attacks from others towards himself.

B. Overall SystemEven though we presented the agents’ minds architecture

and the demonstrator, we still do not know how these twomodules communicate with each other. Thus, in this Sectionwe will describe the overall system. As we can see in Figure4, three main modules compose the overall system: agents’minds, world simulation and realisation engine. We alreadypresented the first and the last modules in the previousSections. However, we still need to describe the worldsimulation, which integrates these two modules.

Figure 4. Dream Theatre demonstrator’s overall system is composed bythree main modules: the agents’ minds, in FAtiMA’s agents architecture,the world Simulator, in ION Framework, and the demonstrator, developedin Unity3D.

1) World Simulation: The world simulation module, aswe stated earlier, functions as the integration layer betweenthe demonstrator and the agents minds. On the other hand,the world simulator, provided by the ION Framework [20],also aims at providing means to simulate dynamic environ-ments.

To achieve this, the ION framework [20] provides themeans to simulate dynamic environments. To do so, itidentifies four basic elements: Entities, Properties, Actions,and Events. Entities populate the simulation universe. Theycan have Properties and change the world through the useof their Actions. As the ION Framework model is based onthe Observer pattern, Entities register to Events. In which,these Events are raised whenever any change to the worldoccurs. For example, if a Property is changed, an Action’s

5

state changes (such as, when it starts or stops), or an Entityenters or leaves the simulation.

2) Integrating Agent’s Minds: FAtiMA agents’ mindsare linked to their respective graphical representations, inUnity3D, through their representations in the ION simulation(see Figure 5). For such, besides the elements we mentionedpreviously in the ION simulation, agents’ minds need anadditional element in the simulation: the Remote Mind.

Figure 5. Communication details between: Unity3D character’s scriptcomponents and ION simulation elements; and between those elements anda FAtiMA agent. The solid arrows define communication through functions,while the dashed arrows signify communication through socket messages.

A script in Unity3D launches the FAtiMA agent’s mindprocess and connects it with the correspondent representa-tion, in the ION simulation. For that, the FAtiMaMind scriptcreates the Remote Mind element in the ION simulation,which registers to all the world events, such as, whenan Entity enters or leaves the simulation, an Action beingperformed, or a Property value changed. With that, theRemote Mind reports all these Events to the FAtiMA agentmind (see Figure 5).

In order to communicate with the FAtiMA agent’s mind,the Remote Mind establishes a socket communication withit. Therefore, all the registered events that occur in thesimulation are transformed into FAtiMA events and then sentto the agent’s mind.

On the other hand, the actions that the agent will perform(by action tendencies or deliberated actions) are sent to theRemoteMind (see Figure 5). These actions are parsed andgenerated into Action start requests to the simulation.

C. Illustrative Example

In order to illustrate how Dream Theater works, let’sanalyse a small situation.

In the first week of playing, the user is faced with aplay that only needs two characters. He has Andy andBob to chose from, but both share the desire for the samerole, the “Hero” role. However, the player can only chooseone of them to have it. Bob has higher proficiency thanAndy, thus he is the most suitable choice for this role.Yet, when Bob receives the role, he appraises this event

and identifies it as an event that interferes positively to hisinterests, thus generating positive emotions. However, whenAndy appraises this event, he will identify it as an event thatinterferes negatively to his self-interests and thus generate anegative emotional state. Further, this is aggravated by thefact that Andy considers the role highly important to him andBob is not his friend. So, with the emotions fired, Andy willget upset enough to approach Bob aggressively, by verballyinsulting him. This process then escalated, and Bob startsfeeling upset, but, given that he was given his preferred role,he limits himself to only question the reason of the insulttrying to resolve this situation. However, Andy appraisesthe situation again negatively, disapproving Bob’s approach,thus getting even more upset. These emotions triggered willlead Andy to plan even more aggressive behaviours towardsBob. As the situation gets even more intense, Bob eventuallyalso reaches a high anger situation, which will lead Bob toplan a way internally how to cope with such pressure. In theend, if not dealt with, Bob ends up giving up the Hero role,thus making the whole cast selection by the child a failure.

VI. EVALUATION

This research aims at developing a prototype conflictmodel to be integrated into an adaptive serious game forteaching resolution skills to children. The model is basedon an emotional system (FAtiMA [11]) and we claim thatthis emotional process is essential to convey concepts asconflict emergence and escalation. Therefore, we conductedan between groups experiment to assess our hypothesis. Forthat, we intended to compare our conflict model, wish wewill emotionally-driven or full-model (FM) condition, with acontrol version where the emotional process was deactivatedon agents, we will name it non-emotional or simplifiedmodel (SM) condition.

A. Measures

With this final evaluation we wanted to measure: (1)whether the participants had the perception of a conflictsituation and the actors’ behaviours towards it; (2) if theusers perceived the influence of emotions on the charac-ters’ behaviours towards the conflict and its influence onthe conflict escalation; and (3) participants’ perceptions ofthe conflict end-state, if characters’ emotionally orientedbehaviours were reasonable and if there is a chance ofrenewable conflict.

B. Evaluation Procedure and Method

A total of 80 participants (19 females, 61 males aged 14-48) took part in the study, which was available through anonline questionnaire. The scenario for evaluation was theDream Theater and the illustrative example on Section V-Cwas the base of our study.

The participants had to watch a video of a user interactingwith (Dream Theater). Similar to the example presented inSection V-C, the user gives the role to Bob, who has ahigher proficiency level and seems to be the natural choice

6

for the given situation, instead of Andy. From this action, asequence of interactions occurs between Andy and Bob.

After watching the video, participants had to answer a setof questions, regarding what they saw on the video. The setof questions were adapted from a self-serving questionnaireon conflict behaviour and escalation [21]. Hence, our ques-tionnaire is divided in three main set of questions. In thesesets, a 5-point likert scale was provided to allow participantsclassify each sentence. Note that participants were blind tothe hypotheses.

1) 1st Set – Characters’ behaviours measurement: First,participants were asked to rate the characters’ behavioursthrough 5-point likert scales. More precisely, they had toindicate to what extent each character’s behaviours was [21]:(a) hostile or friendly (1 - very hostile to 5 - very friendly);(b) competitive or collaborative (1 - very competitive to 5 -very collaborative), (c) frustrating or stimulating (1 - veryfrustrating to 5 - very stimulating), and (d) evil-minded orgood-hearted (1 - very evil-minded to 5 - very good-hearted).

2) 2nd Set – Escalation measurement: After that, asecond set of questions asked participants to rate the processof escalation of the interaction. The questions regarded ifparticipants acknowledged escalation in terms of [21]: (a)to what extent did parties obstruct each other, (b) to whatextent did frustration increase, (c) to what extent did theatmosphere worsen; and de-escalation: (d) to what extentdid parties work out an ideal solution, (e) to what extentdid parties come close, and (f) to what extent were ideasexplored. Each question was rated from 1 - not at all to 5 -very much.

3) 3rd Set – End-state measurement: Finally, the thirdset of questions probed participants about the end-state ofthe situation. Due to the presented scenario, we maintainedonly one of the questions presented in the self-servingquestionnaire [21]: if there is a chance of a renewed conflictbetween the characters. Further, for exploratory reason wealso assessed participants’ perception of the reasonablenessof characters’ behaviours and if participants believe thatsuch behaviours need improvements. These measures wererated by participants through 5-point likert scales, from −2- totally disagree to 2 - totally agree.

C. ResultsThe data was analysed using the Mann-Whitney test to

compare the difference between the two conditions (agentswith the model’s FM version and agents with the model’sSM version). Each measure reported the following results.

Does the user have a perception of a conflict interac-tion?

As we explained previously, this Set of questions ratesthe agents’ behaviours towards the situation. To begin with,questions Q5 to Q8 correspond to the (a) to (d) variablesof the conflict behaviour measures (p. 7) directed to Andy’sbehaviour. After that, questions Q5 to Q8 relates to the samevariables addressed to Bob’s behaviour.

First, we address Andy’s behaviour. Table IV summarisesthe results obtained for the participants rating of Andy’s

behaviour and presents the differences between the twoconditions (FM and SM).

Question1st Set Andy’s Descriptive Statistics

Mann-WhitneyBehaviour Measure

FM (N = 40) SM (N = 40)Mdn[Quartiles] Mdn[Quartiles]

Q5(a) hostile

1[1, 2] 3[2, 3]

U = 298.500

or friendlyp < 0.001r = −0.565

Q6(b) competitive

1[1, 2] 2[2, 3]

U = 455.500

or collaborativep < 0.001r = −0.390

Q7(c) frustrating

2[1, 3] 2[2, 3]

U = 647.000

or stimulatingns

Q8(d) evil-minded

1[1, 2.5] 3[2, 3]

U = 365.500

(or good-heartedp < 0.001−0.488

Table IVMANN-WHITNEY STATISTICS FOR PARTICIPANTS’ PERCEPTION OF

ANDY’S BEHAVIOUR WITHIN THE TWO CONDITIONS (FM AND SM).

Taking a look at the results, from Q5, users’ consid-ered Andy’s behaviour in the FM condition significantly(p < 0.001) more hostile (Mdn = 1.5), in comparison withthe result from the SM condition (Mdn = 3). The samehappened in Q7 and Q8, where his behaviour was foundsignificantly p < 0.001 more competitive (Mdn = 1) andevil-minded (Mdn = 1) in the FM condition, comparingto the SM condition (Mdn = 2 and Mdn = 3). However,Q7 there were no significant differences in the participants’perception of Andy’s frustrating behaviour between the FMcondition and SM conditions.

Despite that, the obtained results are consistent with itsinternal drives to follow a destructive path in a conflict inter-action, modelled in the FM condition. In the SM condition,Andy’s neutral attitude contributed for results distributed inthe middle of the scale.

On the other hand, Table V specifies the results gath-ered for participants’ perception of Bob’s behaviour and itpresents the difference between the FM and SM conditions.

Question2nd Set - Bob’s Descriptive Statistics

Mann-WhitneyBehaviour Measure

FM (N = 40) SM (N = 40)Mdn[Quartiles] Mdn[Quartiles]

Q9(a) hostile

4[3, 4] 3[2, 3]

U = 375.500

or friendlyp < 0.001r = −0.470

Q10(b) competitive

4[3, 4] 2[1, 2]

U = 287, 000

or collaborativep < 0.001r = −0.566

Q11(c) frustrating

3[2, 3] 3[2, 3]

U = 793.500

or stimulatingns

Q12(d) evil-minded

4[3, 4.5] 3[2, 3]

U = 428.000

(or good-heartedp < 0.001r = −0.421

Table VMANN-WHITNEY STATISTICS FOR PARTICIPANTS’ PERCEPTION OFBOB’S BEHAVIOUR WITHIN THE TWO CONDITIONS (FM AND SM).

Bob’s attitude was considered to be more constructive.In Q9, Q10 and Q12 of the FM condition, selected valuesgo towards the other extreme of the scale (Mdn = 4,Mdn = 4, Mdn = 4), compared to Andy’s score. Further,Bob was considered significantly (p < 0.001) more friendlythan the SM condition (Mdn = 3, Mdn = 2, Mdn = 3).

7

Despite the significant (p < 0.001) differences betweenthe two conditions, the non-emotional condition (SM) hasprovoked more responses towards the destructive side ofthe spectrum. We believe that such responses were dueto participants’ subjective view on Bob’s neutral posturetowards the conflict situation. On the SM condition, incontrast to FM, Bob did not give in to Andy’s actions.Therefore, participants might have perceived this as beingcompetitive, as we can see in Q10 (Mdn = 2) of theSM condition, where participants have reported a morecompetitive behaviour in Bob compared to the FM condition(Mdn = 4).

Nevertheless, the results are consistent with Bob’s initiallycooperative behaviour modelled in the FM condition.

Does the user recognise that the conflict escalates?This Set of questions assesses the escalation (questions

Q13 to Q17) and the de-escalation (questions Q18 to Q21)processes of the conflict situation. First, questions Q13 andQ14 are adapted to the context of the scenario to determinethe (a) variable from the conflict escalation measurement(p. VI-B2), where Q13 regards to Andy frustrating Bob andQ14 regards Bob frustrating Andy. Similarly, Q15 and Q16are also adapted in order to assess the (b) variable, regardingAndy and Bob’s increase of frustration, respectively. Finally,questions Q17 to Q21 relate to the variables (c) to (f).

Starting with the escalation process measurement, TableVI contains the results and it compares condition FM withcondition SM.

Question3rd Set Descriptive Statistics

Mann-WhitneyEscalation Measure

FM (N = 40) SM (N = 40)Mdn[Quartiles] Mdn[Quartiles]

Q13(a) To what extent did

5[4, 5] 1[1, 2]

U = 109.500

Andy obstruct Bob p < 0.001r = −0.766

Q14(a) To what extent did

2[1, 2] 4[2.5, 5]

U = 284.000

Bob obstruct Andy p < 0.001r = −0.572

Q15(b) To what extent did Andy

3[2, 4.5] 4[3, 5]

U = 623.000

become more frustrated ns

Q16(b) To what extent did Bob

4[3, 5] 1[1, 2.5]

U = 232.000

become more frustrated p < 0.001r = −0.627

Q17(c) To what extent did

5[4, 5] 4[3, 4.5]

U = 368.500

the atmosphere worsen p < 0.001r = −0.503

Table VIMANN-WHITNEY STATISTICS FOR PARTICIPANTS’ PERCEPTION OF

CONFLICT ESCALATION WITHIN THE TWO CONDITIONS (FM AND SM).

As expected, the results obtained from this set of questionswere quite expressive. First, in question Q13, participantssignificantly (p < 0.001) reported that Andy frustrated Bobmore in the FM condition (Mdn = 5), compared to the SMcondition (Mdn = 1). In contrast, in question Q14, Bob wassignificantly (p < 0.001) rated as less frustrating in the FMcondition (Mdn = 2), then in the SM condition (Mdn = 4).

Next, in Q15, no significant difference was detectedbetween conditions regarding Andy’s increase of frustration.This was quite expected given that in both conditions, Andycould not achieve the role. However, in the FM condition, itwas not expected so many ratings in the middle of the scale(Mdn[Quartiles] = 3[2, 4.5]). We believe that these ratings

were due to participants’ interpretation of Andy’s gleefulsmile, as this emotional expression may have conveyed toparticipants that he achieved his goals.

On the other hand, in Q16, the significant (p < 0.001)difference between Bob’s increase of frustration in the FM(Mdn = 4) and the SM (Mdn = 1) conditions was aspredicted. Since Bob opted out of performing the role in theFM condition, due to reaching the limits of his emotionalcontrol, participants could easily perceive an increase inhis level of frustration. In contrast, Bob’s neutral posturethroughout the SM condition demonstrated no signs offrustration increase.

Finally, as we forecasted, in Q17, participants’ ratings ofthe atmosphere worsening were significantly (p < 0.001)different between the FM (Mdn = 5) and the SM (Mdn =4) conditions. This result is consistent with our process ofconflict escalation which emerges from the emotional statesdeterioration.

Further, considered Q18 to Q21, which assessed conflictde-escalation, no significant differences where found be-tween test conditions, given that no de-escalation processeswere portrayed in both conditions. In both conditions, Q18to Q21 had the same ratings (Mdn = 1).

Does the user recognise a chance for future conflictand characters’ behaviour reasonableness?

The final Set of questions (Q22 to Q26) presented in ourquestionnaire aimed at measuring participants’ perceptionof the conflict end-state. The first two questions from thisSet (Q22 and Q23) assessed participants’ belief of thereasonableness of characters’ behaviours. After that (Q24and Q25), participants stated if they believe that characters’should improve their behaviours. Finally, it was asked forparticipants to rate their belief on the chance of renewedconflict between characters (Q26).

For these measures, Table VII summarises the results ofparticipants’ ratings.

Question 4th Set - End-state MeasureDescriptive Statistics

Mann-WhitneyFM (N = 40) SM (N = 40)Mdn[Quartiles] Mdn[Quartiles]

Q22Andy’s behaviour

−2[−2,−2] 0[−1, 1]

U = 297.000

was reasonable p < 0.001r = −0.585

Q23Bob’s behaviour

0[0, 1] 0[−1, 1]

U = 662.500

was reasonable ns

Q24Andy needs to improve his

2[2, 2] 1[0.5, 2]

U = 389.000

behaviour towards this situations p < 0.001r = −0.505

Q25Bob needs to improve his

1[−0.5, 2] 1[0, 2]

U = 737.000

behaviour towards this situations ns

Q26There is a chance for Andy

2[1, 2] 1.5[1, 2]

U = 724.500

and Bob to discuss again ns

Table VIIMANN-WHITNEY STATISTICS FOR PARTICIPANTS’ FINAL REMARKS ON

THE CONFLICT SITUATION WITHIN THE TWO CONDITIONS (FM ANDSM).

To begin with, in Q22, the results on participants ratingof the reasonableness of Andy’s behaviour was significantly(p < 0.001) different between FM (Mdn = −2) andSM (Mdn = 0) conditions. The result of this measurein the FM condition (Mdn = −2) demonstrated that

8

our emotionally-driven conflict model was able to portraybehaviours destructive enough for participants to rate themas not being reasonable, compared to “neutral” behaviourrepresented in the SM condition (Mdn = 0).

On the other hand, Bob’s ratings of reasonable behaviourwere not as conclusive as Andy’s ratings. No significantdifference was found in Q23 between the FM and SMconditions (Mdn = 0 and Mdn = 0). We believe that suchresults may be due to the duality of Bob’s behaviours inthe FM condition. Given that in the FM condition, initially,Bob’s behaviours towards Andy’s increasingly aggressiveattacks might be rather constructive, as the conflict escalatesto an extreme, Bob starts to lose some emotional control andcopes with this by following a more evasive behaviour (seep. 3). Therefore, we assume that this duality lead participantsto rate Andy’s reasonable behaviour in the middle of thescale, in the FM condition. Further, in the SM conditionthe ratings in the middle of the scale are due to Bob’s“neutrality” throughout the situation, in the SM conditions.

Moreover, in Q24, participants pointed that Andy’s be-haviour in the FM condition needed improvement. A signif-icant (p < 0.001) difference was found between this result inthe FM condition (Mdn = 2) and SM condition (Mdn = 1).This result is coherent with the result of Andy’s reasonablebehaviour rating.

Regarding Bob, in Q25, the results were not conclusiveconcerning participants’ view on behavioural improvement.There was no significant difference between FM (Mdn =1) and SM (Mdn = 1) conditions. Similar to our beliefsfor Bob’s rating of reasonable behaviour (Q23), we positthat the result in the FM condition is due to the dualityin Bob’s behaviours. Therefore, the result for Q24 in theFM condition had a wide discrepancy (Mdn[Quartiles] =1[−0.5, 2]).

Finally, in Q26, participants’ ratings on the possibility ofa renewed conflict situation did not provide a significantdifference between the FM (Mdn = 2) and the SM (Mdn =1.5) conditions. This result was quite unexpected. Taking alook at each condition, even though the result for the FMcondition was as expected, the SM condition also had thesame score. Hence, we have to discard the argument thatparticipants rated that there was a high probable chance ofrenewed conflict due to the conflict escalation portrayed inthe FM condition.

VII. CONCLUSIONS

As we argued in the beginning of this document, conflictis natural to human life, as it is present in our everydaybasis [22]. However, conflict is mostly seen as a negativeact that needs to be eradicated. Contrarily, researchers haveproven that embracing conflict in a constructive way servesas a tool for constructive personal growth [22], especiallyin young ages [23]. For that, conflict resolution educationhas been applied in schools in order to teach students aboutconflict and how to handle it constructively [24].

Concerning education, new teaching tools have been re-searched, such as, serious games. This kind of games have

been used for raising awareness [4], teaching about subjects[2], how to deal with bullying [3], and others.

Motivated by this, the SIREN project intends to create aseries of conflict resolution games to teach children on howto deal with conflict situations [25]. Our contribution for thisproject concerns the creation of a conflict resolution gamethat is populated with autonomous virtual characters.

In order to achieve our hypothesis, we investigated theoryof conflict in social psychology. From this we demonstratedspecial regard to Thomas’ [6] definition of conflict as “theprocess which begins when one party perceives that anotherhas frustrated, or is about to frustrate, some concern of his”and Castelfranchi’s [26] definition of “full social conflict”,which exists when “there is the subjective awareness of thecompetitive situation”.

Further, we reviewed educational games for raising aware-ness on general populate [4], teaching children about asubject [2] and games that involve conflict in the matterof: social interactions [27], intractable conflicts [28], andbullying [3]. However, from the reviewed games, onlyFearNot! [3] assessed emotion in the conflict situations,which is supported by FAtiMA architecture of emotionalagents [11].

With the conflict theory review, we implemented anagents’ model that comprises emotional affect in the conflicthandling process.

To demonstrate our model, we implemented a conflict res-olution game’s demonstrator, Dream Theatre. This demon-strator was populated with synthetic characters modelledby our conflict model in an emotional agents’ architecture,FAtiMA [11].

In order to prove our hypothesis, we performed a betweengroups experiment with 80 participants. Each participant wasrandomly assigned to one of two test conditions. One condi-tion supported our conflict model, where agents’ behavioursare emotionally-driven, we called it emotionally driven orfull-model (FM) condition. Another condition, which servedas a control condition to test our hypothesis, imbued agentswith a variation of our model where the emotional processwas deactivated, we called it non-emotionally driven orsimplified-model (SM) condition.

With the experiment’s results, we compared both con-ditions to assess if our model’s condition (FM) issuedbetter perceptions of the conflict situation, compared to thesimplified-model (SM) condition. From the result’s analysis,we confirmed that in the FM condition, participants hada better perception of the characters’ behaviours towardsconflict, and their roles in the situation. Further, in the FMcondition, the characters’ escalating behaviours influencedby their emotional state conveyed the clear perception ofconflict escalation. These results where consist with yourhypothesis.

REFERENCES

[1] C. Morgado and I. Oliveira, “Mediacao em contexto escolar:transformar o conflito em oportunidade,” Exedra, vol. Volume1, pp. 43–56, June 2009.

9

[2] J. P. Rowe, B. W. Mott, S. W. McQuiggan, J. L. Robinson,S. Lee, and J. C. Lester, “Crystal island: A narrative-centeredlearning environment for eighth grade microbiology,” 2009.

[3] R. Aylett, M. Vala, P. Sequeira, and A. Paiva, “Fearnot! –an emergent narrative approach to virtual dramas for anti-bullying education,” in Virtual Storytelling. Using VirtualReality Technologies for Storytelling, ser. Lecture Notesin Computer Science, M. Cavazza and S. Donikian, Eds.Springer Berlin / Heidelberg, 2007, vol. 4871, pp. 202–205.

[4] G. Rebolledo-Mendez, K. Avramides, S. de Freitas, andK. Memarzia, “Societal impact of a serious game on raisingpublic awareness: the case of floodsim,” in Sandbox ’09Proceedings of the 2009 ACM SIGGRAPH Symposium onVideo Games. ACM New York, NY, USA, 2009, pp. 15–22.

[5] G. Yannakakis, J. Togelius, R. Khaled, A. Jhala, K. Kar-pouzis, A. Paiva, and A. Vasalou, “Siren: Towards adaptiveserious games for teaching conflict resolution,” in 4th Eu-ropean Conference on Games Based Learning (ECGBL10),Copenhagen, Denmark, 21-22 October 2010, 2010.

[6] K. W. Thomas, “Conflict and conflict management: Re-flections and update,” Journal of Organizational Behavior,vol. 13, pp. 265–274, 1992.

[7] B. Laursen and C. A. Hafen, “Future directions in the studyof close relationships: Conflict is bad (except when it’s not),”Social Development, vol. 19, no. 4, pp. 858–872, November2010.

[8] D. G. Pruitt, The Handbook of Conflict Resolution: Theoryand Practice, 2nd ed. Jossey-Bass, 2006, ch. Some ResearchFrontiers in the Study of Conflict Resolution, pp. 849–880.

[9] L. Kriesberg, Conflicts: From Escalation to Resolution,3rd ed. Rowman & Littlefield Publishers Inc, 1998, ch.Analyzing Social Conflicts, pp. 1–26.

[10] E. G. Lindner, The Handbook of Conflict Resolution: Theoryand Practice, 2nd ed. Jossey-Bass, 2006, ch. Emotion andConflict, pp. 268–311.

[11] J. Dias and A. Paiva, “Feeling and reasoning: A computationalmodel for emotional characters,” in Progress in Artificial In-telligence, ser. Lecture Notes in Computer Science, C. Bento,A. Cardoso, and G. Dias, Eds. Springer Berlin / Heidelberg,2005, vol. 3808, pp. 127–140.

[12] C. Tessier, H.-J. Muller, H. Fiorino, and L. Chaudron, “Agentsconflicts: New issues,” in Conflicting Agents, ser. MultiagentSystems, Artificial Societies, and Simulated Organizations,C. Tessier, L. Chaudron, H.-J. Muller, and G. Weiss, Eds.Springer US, 2002, vol. 1, pp. 1–30.

[13] M. Maiese, “Emotions,” in Beyond Intractability, G. Burgessand H. Burgess, Eds. University of Colorado,Boulder: Conflict Information Consortium, July 2005,http://www.beyondintractability.org/bi-essay/emotion Linkactive in May 2012.

[14] E. Raider, S. Coleman, and J. Gerson, “Teaching conflictresolution skills in a workshop,” in Handbook of ConflictResolution, 2nd ed. Jossey-Bass, 2006, ch. 31.

[15] H. Xie, D. J. Swift, B. D. Cairns, and R. B. Cairns, “Ag-gressive behaviors in social interaction and developmentaladaptation: A narrative analysis of interpersonal conflictsduring early adolescence,” Social Development, vol. 11, no. 2,pp. 205–224, 2002.

[16] C. Bell and F. Song, “Emotions in the conflict process: anapplication to the cognitive appraisal model of emotionsto conflict management,” International Journal of ConflictManagement, vol. 16, no. 1, pp. 30 – 54, 2005.

[17] J. Dias, I. Leite, M. Kriegel, M. Y. Lim, C. Martinho, S. Mas-carenhas, A. Paiva, A. Pereira, R. Prada, and A. Wichert,“Architecture development with action-selection mechanism,”2010, unpublished Deliverable 5.3.

[18] A. Ortony, G. L. Clore, and A. Collins, The CognitiveStructure of Emotions. Cambridge University Press, July1988.

[19] A. L. Miller and S. L. Olson, “Emotional expressivenessduring peer conflicts: A predictor of social maladjustmentamong high-risk preschoolers,” Journal of Abnormal ChildPsychology, vol. 28, pp. 339–352, 2000.

[20] M. Vala, G. Raimundo, P. Sequeira, P. Cuba, R. Prada,C. Martinho, and A. Paiva, “Ion framework – a simulationenvironment for worlds with virtual agents,” in 9th Interna-tional Conference on Intelligent Virtual Agents - IVA2009,ser. Lecture Notes in Computer Science, vol. 5773 / 2009.Amsterdam, The Netherlands: Springer Berlin - Heidelberg,September 2009, pp. 418 – 424.

[21] C. K. W. De Dreu, A. Nauta, and E. Van de Vliert, “Self-serving evaluations of conflict behavior and escalation of thedispute,” Journal of Applied Social Psychology, vol. 25, pp.2049—-2066, 1995.

[22] M. J. M. C. Luıs, “A assembleia de turma como dispositivode mediacao de conflitos,” Master’s thesis, Universidade deLisboa, Faculdade de Psicologia e de Ciencias da Educacao,2009.

[23] W. J. Kreidler, Teaching Conflict resolution through children’sliterature (Grades K-2). Scholastic Inc., 1994.

[24] D. W. Johnson and R. T. Johnson, “Conflict resolutionand peer mediation programs in elementary and secondaryschools: A review of the research,” Review of EducationalResearch, vol. 66, no. 4, pp. 459–506, Winter 1996.

[25] Y.-G. Cheong, R. Khaled, C. Grappiolo, J. Campos, C. Mart-inho, G. P. D. Ingram, A. Paiva, and G. Yannakakis, “A com-putational approach towards conflict resolution for seriousgames,” in Proceedings of the Sixth International Conferenceon the Foundations of Digital Games, ser. FDG ’11. ACM,2011.

[26] C. Castelfranchi, Conflict Ontology. Springer-Verlag NewYork, Inc., 2000, pp. 21–40.

[27] J. McCoy, M. Treanor, B. Samuel, B. Tearse, M. Mateas,and N. Wardrip-Fruin, “Comme il faut 2: A fully realizedmodel for socially-oriented gameplay,” in Proceedings of theIntelligent Narrative Technologies III Workshop. ACM NewYork, NY, USA, 2010.

[28] T. Buch and S. Egenfeldt-Nielsen, “The learning effect ofglobal conflicts: Palestine,” Media@Terra conference, 2006.

10