Modelling of Personality in Agents: FromPsychology to Implementation

Sebastian Ahrndt, Johannes Fahndrich, and Sahin Albayrak

DAI-Laboratory, Technische Universitat Berlin,Faculty of Electrical Engineering and Computer Science,

Ernst-Reuter-Platz 7, 10587 Berlin, Germanysebastian.ahrndt@dai-labor.de (Corresponding author)

Abstract. There is increasing interest in the agent community to in-tegrate the concept of emotions and artificial agents. The spectrum ofavailable solutions reaches from applications and models of emotions tocomplete axiomatised logics. Despite the rich offer of solutions, availableworks neglect individual personality as a significant factor for the out-come of emotional behaviour pattern. However, different personalities af-fect all relevant phases of human decision-making processes. Hence, thispaper introduces and discusses existing personality theories and high-lights the fact that one of them is widely accepted in psychology andshould be adopted by the agent-community. We integrate the charac-teristics of this personality theory into the life-cycle of BDI agents anddiscuss two different versions of the BDI algorithm – a naive one and onethat balances the commitment between means and ends. The outlinedalgorithm is implemented as a prototype model in AntMe!, an agent-based simulation environment for behavioural studies. The experimentsperformed in this environment show that personality indeed affects allrelevant phases of the decision-making process, laying the foundationsfor future empirical studies.

1 Introduction

Over the last years, the agent community presented several approaches to bringemotions to artificial agents. Available solutions reach from modelling and apply-ing emotions [18, 19, 21, 24] to (complete axiomatised) logics of emotions [1, 12,26, 31]. The latter provide discussions about the effects of emotions on decision-making in a use-case independent and principle manner.

Comparing this rich offer of works with the available literature on agents in-corporating another important aspect of human behaviour reveals an interestinggap: The missing link between personality and emotions in agent-based systems.But is it not that our personality affects our emotions and determines our entirebehaviour? Following Ozer and Benet-Martınez [25], personality, in fact, is asignificant factor for human behaviour and determines the individual outcomeof essential behavioural processes, e.g. cognition and emotional reactions.

1.1 Towards Human Personality

At this point, several questions arise where the first one addresses the conceptof personality itself. In psychology different theories exist that explain the be-haviour of humans describing the humans personality along personality traitsor types. These theories have in common that each trait/type is a characteristicfeature of a human, which can be used to explain the humans behaviour and itsmotives along patterns of behaviour. Nowadays two big theories about humanpersonality exist [17]: The Five-Factor Model of personality [22] (FFM) and theMyers-Briggs Type Indicator [23] (MBTI). Both theories and the differences be-tween them are discussed in detail in a prior work [2] from the psychologicalperspective. The argumentation comprises the origin (empirical vs. theoretical),the scale types (traits with continuous scale vs. types representing clusters), thecompleteness (both incomplete), and the reliability or consistency of assessmentsaccording to both theories. Balancing the arguments, we concluded that psychol-ogists tend to accept FFM as a conceptual framework. For the agent-community,this would imply to apply the FFM for experiments using personality traits, e.g.when modelling human-behaviour for virtual humans.

FFM introduces—as indicated by the name—five dimensions characterisingan individual. These are openness to experience, which is related to a person’spreference to act inventive, emotional and curious vs. acting consistent, conser-vative and cautious; conscientiousness, which is related to a person’s preferenceto act efficient, planned and organised vs. acting easy-going, spontaneous andcareless; extraversion, which is related to a person’s preference to act outgoing,action-oriented and energetic vs. acting solitary, inward and reserved; agreeable-ness, which is related to a person’s preference to act friendly, cooperative andcompassionate vs. acting analytical, antagonistic and detached; and neuroticismwhich is related to a person’s preference to act sensitive, pessimistic and nervousvs. acting secure, emotionally stable and confident.

1.2 Agents with Personality

In research on agent-based systems, models of human personality are compre-hensively used for the implementation of (microscopic) traffic simulation frame-works [20] and the agent-based simulation/visualisation of groups of people [9,15]. The work of Durupinar et al. [15] shows how the introduction of differentpersonalities into single agents influences the behaviour of a crowd. For thissimulation the authors applied FFM. Other areas include human-machine inter-action [13], in particular conversational agents/virtual humans [4, 16] and life-likecharacters [5]. The latter outlines three projects that apply two dimensions ofFFM (extraversion and aggreeableness). The effects are interpreted in a rule-based or scripted manner. Another branch of research focuses on modelling andexamining the effects of personalities on interactions between agents and theirenvironments. In particular, the effects of personalities in cooperative settings.

Talman et al. [32] present a work that illustrates the use of a rather simpleabstraction of personality types. Personalities of agents are determined through

the two dimensions cooperation and reliability, which are used to measure thehelpfulness of an agent. The agents have to negotiate and cooperate as coopera-tion is an inherent part of the game they play. During repeatedly played gamesthe agents reason about each others helpfulness along the two dimensions. Asan effect they try to respond more effectively by customising their behaviourappropriately for different personalities.

Campos et al. [8] present a work employing MBTI, which is here restricted totwo of its dichotomies. It is integrated into the reasoning process of a BDI agentand the work proves that different personality characteristics lead to varieties inthe decision-making process in a simulation specifically designed for the paper’suse-case.

In an early work, Castelfranchi et al. [10] present a framework to investigatethe effects of personalities on social interactions between agents, such as dele-gation and help. The agents apply opponent modelling in terms of personalitytraits to motivate interactions. However, the work discusses personality traits asan abstract concept without relation to psychological theories.

The work that is most closely related to our work is presented by Salvitand Sklar [29, 30]. That is the case, because the authors setup an experimentvalidating the impact of the MBTI onto the decision-making process of agents.In order to do so, MBTI is integrated into a sense-plan-act structure and thebehaviour of each MBTI type is analysed in a simulation environment calledthe ‘Termite World’. The results underline the hypothesis of the paper that thedifferent personality types act in quite different ways.

Discussion There are several works applying personality to different domainssuch as traffic simulation or the study of cooperative decision making. They fre-quently implement the effects of personalities specifically for the single use-case,without discussing/evaluating how this can be done in a more generic manner1.Although this is a prerequisite to close the gap between emotional agents andagents with personality there are even some works that slightly touch the topic ofpersonality when working on emotional agents (cf. [7, 19, 24]). Thus approachesare discussing architectural considerations from the software engineering per-spective. However, the literature overview also shows that the majority of worksuses simplified models of personality that are not based on psychology findingsor apply specific traits. Stunningly, reasons for using either the FFM or MBTIare missing in all considered works.

1.3 Motivation and Problem

To conclude, the questions of how to incorporate personality into agents was notsatisfiable approached by agent researchers, yet. In fact, influences of personal-ity on the decision-making process of agents were discussed in the early work

1 Except the work of Salvit and Sklar, which was presented in this workshop series atthe 2nd HAIDM in 2012.

of Castelfranchi et al. [10] as an abstract concept without a relation to psycho-logical findings. The work of Salvit and Sklar [29, 30] discussed and evaluatedthese influences with respect to the MBTI and concluded ‘that some agent per-sonality types are better suited to particular tasks—the same observation thatpsychologists make about humans’ [30, p. 147]. As psychologist tend to acceptthe FFM as personality framework and at the same time tend to refuse theMBTI, the motivation for this work is to confirm the findings of Salvit and Sklarwith respect to the FFM. This will finally provide the foundation for the logicalformalisation of personality influences on an agents decision making process asclaimed as prerequisite to close the gap between emotional agents and agentswith personality elsewhere [3].

Remainder In the following, we present an extension of a selected BDI algo-rithm as well as particulars about the implementation of this algorithm in amulti-agent simulation in Section 2. In doing so, we assess the level to which per-sonality affects the different stages of the sense-plan-act lifecycle. We establishthis assessment by means of simulation results in Section 3. Collected simulationresults show that the quality by which problems are solved indeed varies with theproblem-solvers personality, that is, problem solving can be altered (and some-what improved) by a careful personality-specific task assignment. We concludeour findings in Section 4.

2 Modelling Agents with the FFM

In the following, we will discuss how the FFM can be integrated into the BDImodel of agency [28], a popular model for the conceptualisation of human be-haviour. BDI agents separate the current execution of a plan from the activityof selecting a plan using the three mental concepts belief, desire and intention.The life-cycle of a BDI agent comprises four phases [34, pp. 23–32], namely theBelief Revision, the Option Generation, the Filter Process, and the Actuation.In our model, the phases of the BDI cycle are influenced by the characteris-tics of a personality in different ways. For instance, the trait conscientiousnessstrongly influences the goal-driven behaviour of an agent, whereas the trait ex-traversion influences the agent’s preference to interact with others. Table 1 liststhe influences of the different characteristics of FFM on the different phases ofthe BDI life-cycle. These influences address the intensity by which one person-ality trait influences a phase and thus (only) highlights the traits that are mostinfluential. Indeed, this classification is discussable as it reflects our own inter-pretation of the FFM traits in comparison with the BDI phases. To substantiateour interpretation we took in account works of different authors investigatingthe relation between personalities and behaviour types (e.g., [14, 27]). Further-more, we learned about the influences by experiments that provide findings forthe relation between personalities and specific stages of the decision cycle (e.g.,effects on coping strategies [11], effects on information processing [6]).

Table 1: In order not to value the influence in terms of being negative or positive,the list only highlights the traits that are most influential in each phase.

O C E A N

Belief Revision × ×Option Generation × × ×Filter Process × × × × ×Actuation × × ×

2.1 Personality and the naive BDI lifecycle

To explain the model, we will use the following syntax introduced by MichaelWooldridge [34, pp. 69–90] for the ‘Logic Of Rational Agents’ (LORA):

– P : Per is the collection of personality traits of the agent;– ρ : Percepts is the information that the agent perceives/receives in its envi-

ronment;– B : ℘(Bel), D : ℘(Des), I : ℘(Int) are the sets of beliefs, desires and

intentions, respectively;– π : Act∗ representing the current sequence of actions taken from the set of

plans over the set of actions DAc, i.e. the current plan; and– α : Act representing the action that is executed.

Algorithm 1 shows an adapted BDI life-cycle that involves personality as influ-ence during the different stages. All personality traits are considered during theprocess. Furthermore, we assume that the personality does not change duringthe life-cycle of an agent. That is based on the finding that we as humans havea stable personality over our lifespan as adults [33].

The cycle starts with the perception of information. During this stage theagent receives new information from the environment (Env) using its sensors,which also comprises messages (Msg) from other agents (communication acts).The perception is not affected by the personality, as humans are not able torestrict their perception during the cognition. This is a deliberate process takingplace in the next step of the cycle. Formally, the signature of the perceptionfunction percept is defined as:

percept : Env ×Msg → Percepts.

The next step of the BDI life-cycle is the Belief Revision. That means thatgiven perceptions (ρ) are computed with respect to the current personality (P )to update the actual beliefs (B). The belief revision function beliefRevision isdefined as:

beliefRevision : ℘(Bel)× Percepts× Per → ℘(Bel).

After this step the set of beliefs can contain information about the environment,the state of the agent itself (e.g., energy level, injuries like sensory malfunctions)

Algorithm 1 A BDI cycle that incorporates personality into the decision mak-ing process.

Input: Binit, Iinit, P ; Output: -

1: B ← Binit, I ← Iinit

2: while true do3: ρ ← percept(Env, Msg)4: B ← beliefRevision(B, ρ, P )5: D ← options(B, I, P )6: I ← filter(B, D, I, P )7: π ← plan(B, I, P )8: while not empty(π) do9: α ← hd(π)

10: execute(α, P )11: π ← tail(π)12: end while13: end while

and facts that were received via communication. In our model the O and Acharacteristics influence this phase most frequently, as they influence the in-terpretation what the new measurement means for the agent and how trustfulthe agent is when receiving information from others. One essential reason todistinguish between perceptions/beliefs derived from the environment and per-ceptions/beliefs derived from other agents is the characteristic of the personalitytrait agreeableness, which indicates the preference to trust others.2 We imple-mented this behaviour (the influence of the trait A during the belief revision) forour simulation environment using the characteristic of the personality trait aslikelihood. For example, an agent with A = 1.0 always trust information receivedvia communication acts, whereas an agent with A = 0.0 always reject them.

The next step is the Option Generation, where the agent generates its desires(D) taking into account the updated beliefs, the currently selected intentions (I)and the personality. The option generation is mainly influenced by the C, A andN characteristics, as these traits indicate the preferences to follow picked goals,the tendency to act selfish or altruistic, and the reaction of the agent to externalinfluences. This deliberation process is represented by the function options withthe following signature:

options : ℘(Bel)× ℘(Int)× Per → ℘(Des).

The generated desires are a set of alternatives (goals) an agents wants to fulfil,which are often mutually exclusive. As the option generation should produceall options available to the agent the influence of the personality is restrictedto the persistence of already selected intentions. Again, we implemented the

2 In fact, it might be hard to clearly distinguish the information sources. That isbecause other agents are part of the environment and the observation of the be-haviour of other agents might thus be both an observation of the environment andan (implicit) communication act.

influence by interpreting the traits as likelihood, e.g. an agent with C = 1.0always maintain an intention as option regardless of the current beliefs aboutthe world.

The third stage is the Filter Process where the agent chooses between com-peting desires and commits to achieve some of them next. The filter processis influenced by the preferences to vary activities over keeping a strict routineand the level of self-discipline (O, C), the need to act in harmony with otheragents (A, N) and even the tendency to generally interact with others (E). Forexample, variations of C influence an agent’s preference to detach the prior se-lected intentions. As another example, variations of A and E influence an agent’spreference to commit to selfish/altruistic goals. The filter function is defined as:

filter : ℘(Bel)× ℘(Des)× ℘(Int)× Per → ℘(Int).

The personality helps to prioritise the different intentions and for example in-dicates to what extent an agent acts goal-driven, prefers interaction, varies theactivities. It selects the best option from the point of view of the agent based onthe current beliefs, with respect to the prior selected option. Again interpretingthe traits as likelihood, the filter process is implemented by, e.g., prioritisingintentions that imply interaction with others using the characteristic of E.

The last stage is the Actuation, in which the agent creates/selects the plan(π) and influences the environment performing actions (α). This phase is mainlyinfluenced by the creativity level of the agent (O), the tendency to apply actionsin a decent manner (C) and the preference to interact with others (E). Theactual plan is then generated for the selected intentions and executed, which isdefined as:

plan : ℘(Bel)× ℘(Int)× Per → Act∗.

The execution of actions as plan-elements directly influences the environmentand the personality indicates how accurate an agent behaves (C). This is arather vague argument for agents. To set an example, imagine a robot thatshould perform a motion from one point to another in a specific time frame. Thelevel of conscientiousness then can be used to implement a noise level added tothe target location or time frame boarders. Indeed, this seems to be curious whenconsidering artificial agents but is one important difference between humans. Theactuation function execute is formally defined as:

execute : Act× Per

2.2 Balancing commitments to means and ends

The prior explained algorithm is one variant of a BDI agent following a blind-commitment strategy and being overcommitted to both, the ends (i.e., the se-lected intentions respectively the world state the agents wants to achieve) andthe means (i.e., the generated plan to achieve the intended world state). Thiscommitment strategy is acceptable for the simulation environment used within

this work as: the domain is tick-based, the plans a rather short, and plans ex-ecuted for an intention are fixed, making the time required to generate a plannegligible.

However, using the provided explanation the algorithm can be adapted toproduce reactive and single- or open-minded behaviour, which might be eitherbold or cautious. Algorithm 2 shows one variant of a BDI life-cycle that is notovercommitted to intentions or plans (adapted from published work [34, pp.31]). In order to achieve these properties the actuation stage is extended with a

Algorithm 2 A BDI cycle that incorporates personality into the decision mak-ing process that is not overcommitted to the means or ends.

Input: Binit, Iinit, P ; Output: -

1: B ← Binit, I ← Iinit

2: while true do3: ρ ← percept(Env, Msg)4: B ← beliefRevision(B, ρ, P )5: D ← options(B, I, P )6: I ← filter(B, D, I, P )7: π ← plan(B, I, P )8: while not (empty(π) or succeeded(I, B) or impossible(I, B)) do9: α ← hd(π)

10: execute(α, P )11: π ← tail(π)12: ρ ← percept(Env, Msg)13: B ← beliefRevision(B, ρ, P )14: if reconsider(I, B) then15: D ← options(B, I, P )16: I ← filter(B, D, I, P )17: end if18: if not sound(π, I, B) then19: π ← plan(B, I, P )20: end if21: end while22: end while

perception and belief revision stage. This is done as each action takes some exe-cution time and thus the environment might change to a state where the currentintention or the current selected plan is not of relevance anymore. The processoutlined in Algorithm 1 can not recognise these facts as it strictly executes aonce selected plan. Also introduced are some methods that help the agent to de-cide whether it must reconsider its current intentions or whether the currentlyselected plan is sound. The inner while condition is further extended with twoconditions, which validate whether the intentions were successfully achieved orbecame impossible.

Taking the above argumentation about the influence of the personalty intoaccount one might argue that the traits must affect the condition checks as well.For example, the trait C as one of the major influences during the execution(remember the noise example) could influence the succeeded check. The trait Nindicating the emotional stability could influence the impossible check, in termsof ‘more rounds, more stressful’. However, the extension mechanism proposedaffects the existing stages of the BDI cycle and we argue that these effects takeplace in these stages. Making the influence in the condition checks redundant.For example, whether an intention was successfully achieved is recognised inthe belief revision and thus will make the effects of C implicit available in thecondition check.

3 Experimental Setup and Results

To evaluate the model we implemented it for the multi-agent simulation environ-ment AntMe!3. The main objective of each ant colony is to collect as much food(apples, sugar) as possible and to defend their own anthill from enemies such asother ant colonies and bugs. Each simulation run encompassed 5000 time-steps,where each ant in each time-step completes the BDI cycle of sensing its envi-ronment, updating its beliefs, desires and intentions and executing. The antsare able to sense their location and to recognise whether or not they are trans-porting food, the location of food, other ants, scent-marks, and enemy withintheir range of sight. The scent-marks are used to determine what other ants ofthe own colony are targeting and to highlight the occurrence of enemies. Thepossible actions are goStraight, goAwayFromPOI, goToPOI, goToNest (‘moveactions’), turnToPOI, turnByAngle, turnAround, turnToGoal (‘turn actions’),pick-up and drop-off food, attack, and put scent-mark. Fig. 1a shows a screen-shot of the simulation environment.

Using the introduced model we expect that the ants’ behaviours vary whenadjusting the personality traits. In particular we expect that an ant populationwith high values in the trait openness (O+) does more exploration more thana population with low values (O-).4 That means that O+ ants are expectedto find sugar and apples earlier. At the same time, we expect the O- ants toharvest sugar faster as a consistent behaviour is favourable for this task, whichincludes walking the same route multiple times.5 We expect that high valuesin the trait conscientiousness (C+) lead to more collected food, as such antswill not drop food when facing other goals such as attacking/running away frombugs. At the same time, we expect low valued ants (C-) to have a lower chanceto starve during the search for food as collecting food is the most importantdesire. Extroverted ants (E+) are expected to communicate more frequentlywith other ants by putting scent-marks as markers for the occurrence of sugar,

3 For further information about the simulation environment the interested reader isreferred to http://www.antme.net/.

4 The −, + label represent a value in the interval [0.0, 0.5), [0.5, 1.0] respectively.5 In other words, openness indicates the choice between exploration and exploitation.

apples and bugs more frequently. However, this effect correlates with the effectof the trait agreeableness, indicating whether an ant trusts information receivedfrom other ants (A+) or not (A-). We expect that high valued ants in both traitscollect food more frequently. The neuroticism trait indicates the ants’ emotionalstability. We expect high valued ants (N+) to avoid dangerous situations suchas bugs and hostile ants – resulting in lower numbers of eaten ants and killedbugs. However, the effect of this trait correlates with the level of trust (A+ vs.A-) and the level of self-discipline (C+ vs. C-).

The upper part of Table 2 shows the correlation matrix for all personalitytraits and the measurable features of an AntMe! simulation. For this we sim-ulated the permutation of the minimum and maximum values for each trait,resulting in 25 = 32 ant populations. The features comprise the collected applesand the collected sugar, the number of eaten and starved ants, and the numberof killed bugs. For each permutation the values were averaged over 50 simulationruns, where each simulation run started with the same point of origin of theant hill, apples, and sugar. Occurrence of bugs is randomised and each deceasedant is instantly replaced with a new one. As indicated in the correlation matrix,the majority of effects that were postulated are observable in the simulation.To start with, the matrix indicates that O+ ants collect less food than O- antsand that this behaviour is most notable for the collected sugar. Still, we pos-tulated that O+ ants will find sugar earlier. This effect can be observed, i.e.when building average about all O+/O- populations the O+ ant populationsstart approximately 2% earlier with the collection, but collect food slower thanO- ant populations.

The lower part of Table 2 lists the results of all ant populations representinga permutation of the minimal and maximum values of the personality traits.It is emphasised that different types of personality lead to different simulationresults. For example, an ant population with maximum values (1,1,1,1,1) collectsmore apples and sugar, kills fewer bugs and loses fewer ants through bugs thanan ant population with minimum values (0,0,0,0,0). Still, for the latter a lowernumber of starved ants can be observed. Here, the traits E and A influence theoccurrence of scent-marks and the interpretation (trust) of the very same thing.The trait C implies that already picked-up food is not dropped through newpercepts as collecting food is the most important goal for the ants. The trait Naffects the fight behaviour of the ants leading to fewer/more eaten ants/killedbugs, respectively.

The effects of the personality traits are also visible in the paths an ant popula-tion takes. Fig. 1b - Fig. 1d are showing the path heatmaps for three populations.They emphasise the effects of the trait O, which affects an ant’s preference of act-ing explorative vs. exploitative or following a conservative vs. curious behaviour(i.e., staying in known areas vs. eager to explore new areas). At the same point,the depiction visualises how cooperative the ants act, visible through the roundartefacts highlighting the occurrence of apples – collecting apples is a cooperativetask.

Table 2: Correlation matrix between measured items and personality traits (up-per part) and collected information for a set of ant populations. Lowest andhighest measured items ar highlighted in bold.

Apple Sugar Eaten Starved Bugs

O -0.068 -0.444 -0.043 -0.209 0.027C 0.545 0.425 -0.454 0.893 -0.027E -0.150 0.072 0.002 -0.119 -0.009A 0.261 0.501 -0.430 0.107 -0.554N 0.305 0.114 -0.436 0.125 -0.554

values below are ordered according to the OCEAN acronym

(0,0,0,0,0) 8.4 18.4 281.6 6.0 2.5(0,0,0,0,1) 19.5 81.0 84.7 130.9 0.0(0,0,0,1,0) 19.8 83.2 82.8 131.5 0.0(0,0,0,1,1) 19.4 78.2 84.7 131.6 0.0(0,0,1,0,0) 8.2 10.6 284.8 3.5 1.8(0,0,1,0,1) 18.6 43.1 97.2 39.8 0.0(0,0,1,1,0) 16.7 113.1 83.0 55.1 0.0(0,0,1,1,1) 16.1 108.3 83.4 55.3 0.0(0,1,0,0,0) 19.0 75.6 117.4 146.9 3.3(0,1,0,0,1) 19.4 88.3 54.5 204.8 0.0(0,1,0,1,0) 19.7 90.1 54.3 203.5 0.0(0,1,0,1,1) 19.3 86.8 55.7 203.1 0.0(0,1,1,0,0) 19.0 52.9 98.3 162.9 2.1(0,1,1,0,1) 19.3 58.0 51.5 204.7 0.0(0,1,1,1,0) 16.5 181.0 65.9 174.6 0.0(0,1,1,1,1) 16.0 175.7 64.2 175.9 0.0(1,0,0,0,0) 8.5 8.1 285.4 0.0 3.0(1,0,0,0,1) 16.2 46.3 77.1 0.0 0.0(1,0,0,1,0) 16.8 48.1 82.7 0.0 0.0(1,0,0,1,1) 16.2 44.5 80.1 0.0 0.0(1,0,1,0,0) 7.9 6.5 283.9 0.1 3.5(1,0,1,0,1) 15.7 28.9 74.0 0.0 0.0(1,0,1,1,0) 16.2 40.1 74.4 0.0 0.0(1,0,1,1,1) 15.8 39.9 75.0 0.0 0.0(1,1,0,0,0) 19.2 70.5 90.7 167.6 1.6(1,1,0,0,1) 19.5 65.0 54.6 193.5 0.0(1,1,0,1,0) 19.6 69.6 54.3 193.4 0.0(1,1,0,1,1) 19.4 66.7 52.1 197.1 0.0(1,1,1,0,0) 18.8 47.2 99.7 153.0 2.7(1,1,1,0,1) 19.1 50.4 52.6 193.3 0.0(1,1,1,1,0) 19.4 78.6 55.5 184.5 0.0(1,1,1,1,1) 19.3 75.8 54.2 188.4 0.0( 12, 12, 12, 12, 12) 9.7 17.1 270.8 19.5 1.5

(a) Map (b) (0,0,0,0,0) (c) ( 12, 12, 12, 12, 12) (d) (1,1,1,1,1)

(e) (0,1,1,1,0): Max.sugar

(f) (1,0,1,0,0): Min.sugar

(g) (0,0,0,1,0): Min.starved

(h) (0,1,0,0,1): Max.starved

Fig. 1: Fig. 1a shows the map used for the single population simulations. Theoccurrence of food (apples in green, sugar in white) and the location of the ant hillare fixed. The ants goal is to collect as much food as possible and not to die eitherby starving or by fighting against bugs (blue). Fig. 1b–1h show the cumulatedpaths of seven ant populations. As the map is fixed a comparable structureoriginates. Still, the effects of exploration vs. exploitation are visible (coveredarea, curious behaviour, broader paths). The artefacts denote the visibility rangeof the ants and the points apples are spawned, giving an indication of the effectsof scent-marks and the trustfulness of the ants.

3.1 Discussion and Implication

Taking these results into account we can state that the parameters we addedto the BDI lifecycle can be interpreted as personality traits and the resultingbehavioural change of the agents can be interpreted as personality. In addition,we have shown that different personality traits affect the result of the simulationand that some personalities are better suited for particular tasks than others.This extends the work of Durupinar et al. [15] to the complete set of person-ality traits available through the Five-Factor Model. We have also learned thatsuch parameters can influence the behaviour of agents in a domain independentmanner and that one challenge is the task-dependent interpretation of the ef-fect of a personality. Finally, the experiment confirmed the finding of Salvit andSklar [30] that the interpretation of the parameters as personality traits resultsin (personality-)consistent behaviour of agents with respect to the Five-FactorModel instead of the MBTI.

The implication is that the task performance of problem-solvers can be im-proved by carefully assigning personality-specific tasks. To show that, we can usethe derived results to determine which population performs more accurate for a

specific objective. In Table 2 the minimal and maximal values that were reachedby the populations for the different measured categories are highlighted. Is theobjective to collect as much sugar as possible the population (0,1,1,1,0) wouldbe the best choice, whereas the population (1,0,1,0,0) would be the worst choice.Fig. 1e and Fig. 1f show the paths walked by the ants of the population thatcollect the most and the least amount of sugar, respectively. Ants with character(1,0,1,0,0) collect also the least amount of apples, but at the same point attackbugs frequently leading to a high amount of eaten ants. Another example is theobjective to let as least ants starve as possible (ants starve if they not rest; restmeans staying in the anthill). The paths of one of the populations that are notstarving is shown in Fig. 1g. Here the concentration on staying in near distanceto the anthill becomes visible. In contrast, Fig. 1h shows the path heatmap ofan ant population where the individuals avoid periods of rest starving whenexploring the map.

3.2 First Empirical Results

Additionally to the set of populations with extreme values, we applied a realisticset of personalities to the simulation environment. Thus personalities were ele-vated from 19 colleagues of our own institute, which were asked to assess theirpersonality using a questionnaire derived from the IPIP6. Table 3 lists the simu-lation results of five of these personalities. Using realistic values leads to resultsthat are not as distinct as for the extreme values. However, the differences arestill visible especially when the personalities are in more distance to each other.

Table 3: Ten real personalities with corresponding simulation results.

OCEAN Apple Sugar Eaten Starved Bugs

(0.85, 0.70, 0.47, 0.34, 0.83) 13.4 25.8 233.6 47.1 0.7(0.49, 0.95, 0.31, 0.70, 0.48) 17.4 49.9 166.1 104.5 1.3(0.63, 0.63, 0.65, 0.71, 0.64) 13.10 26.68 239.18 42.58 0.70(0.81, 0.84, 0.73, 0.60, 0.46) 13.42 27.34 234.26 46.14 0.90(0.59, 0.59, 0.44, 0.64, 0.60) 11.64 23.14 255.48 29.60 1.24(0.59, 0.69, 0.24, 0.43, 0.49) 10.24 19.26 266.32 21.64 1.40(0.95, 0.95, 0.49, 0.83, 0.73) 18.96 60.66 115.66 144.90 0.26(0.70, 0.35, 0.70, 0.66, 0.56) 10.36 17.88 265.76 21.26 1.46(0.95, 0.69, 0.76, 0.90, 0.83) 18.20 51.32 138.18 123.92 0.02(0.71, 0.71, 0.51, 0.64, 0.41) 11.50 21.88 256.84 28.82 0.84

6 IPIP — International Personality Item Pool: A Scientific Collaboratory for the De-velopment of Advanced Measures of Personality and Other Individual Differences— http://ipip.ori.org/. For the experiment the 100-Item Set of IPIP Big-FiveFactor Markers was used.

The intention to use realistic personalities is to compare these simulationresults with the expectation of the participants in an uninformed and informedstage. To do so we explained the simulation environment and the measurableitems to the participants and asked them to formulate their expectations beforethey were informed about their personalities and afterwards. Until now, theanalysis of these results is work in progress.

4 Final Remarks

In this work, we discussed the current state-of-the-art of those agent-based worksthat integrate personality as a factor for agent-based behaviour. We showed thatthere is a gap between the progress that was made for emotional agents and thatthere is a missing link between both (essential) human behavioural processes.Based on this finding, we took the Five-Factor Model of personality into accountand discussed how it can be integrated into the BDI reasoning process. Wedemonstrated the applicability of our approach by means of AntMe!, an agent-based simulation framework, which provides a completely adaptable test-bedfor behavioural studies. Despite the fact that we simulated ants, we were ableto show that personality affects all relevant phases of agent decision-makingprocesses and conclude that personality-specific task assignment can alter and/orimprove the quality in which problems are being solved. Having done that, wewere able to confirm the findings of Salvit and Sklar [29, 30] with respect to theFFM, which is the initial intention to present this work.

It is important to mention, that we presented a stepping-stone rather than aholistic solution. A more comprehensive implementation should address severalissues, most importantly: The impact of one particular personality trait is alwayssubject to the environmental context of the individual. An introverted person,for instance, is usually cautious when meeting other people for the first time, e.g.,when attending a scientific conference. At the same time, the same person mightact rude, when writing emails or chatting to people they never met. Findingconcepts for this particular characteristic of human behaviour and integratingthese concepts with existing emotional agent approaches is an open topic andrequires both, theoretical work and user-studies. In another work [3], we brieflyintroduce the first steps towards a theoretical integration by presenting thoughtson a logical formalisation of the here presented algorithm. The basic idea is tointegrate personality as an own modal connectivity in the BDI life-cycle.

References

1. Adam, C.: Emotions: from psychological theories to lologic formalization and Im-plementation in a BDI agent. Ph.D. thesis, L’ Institute National Polytechnique deToulouse (2007)

2. Ahrndt, S., Aria, A., Fahndrich, J., Albayrak, S.: Ants in the OCEAN: Modulatingagents with personality for planning with humans. In: Bulling, N. (ed.) Multi-AgentSystems (EUMAS 2014), pp. 1–16. Lecture Notes in Artificial Intelligence, Springer(December 2014)

3. Ahrndt, S., Faehndrich, J., Lutzenberger, M., Albayrak, S.: Modelling of person-ality in agents: From psychology to logical formalisation and implementation. In:Bordini, Elkind, Weiss, Yolum (eds.) Proc. of the 14th Int. Conf. on AutonomousAgents and Multiagent Systems (AAMAS 2015), pp. 1–2. IFAAMAS (May 2015)

4. Allbeck, J., Badler, N.: Toward representing agent behavior modified by personalityand emotion. In: Proc. of the Workshop on Embodied Conversational Agents at the1st Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS). ACM(2002)

5. Andre, E., Klesen, M., Gebhard, P., Allen, S., Rist, T.: Integrating models ofpersonality and emotions into lifelike characters. In: Paiva, A., Martinho, C. (eds.)Proc. of the Workshop on Affect in Interactions Towards a new Generation ofInterfaces. pp. 136–149 (2000)

6. Baumert, A., Schmitt, M.: Personality and information processing. European Jour-nal of Personality 26(2), 87–89 (2012)

7. Bevacqua, E., de Sevin, E., Pelachaud, C., McRorie, M., Sneddon, I.: Buildingcredible agents: Behaviour influenced by personality and emotional traits. In: Levy,P., Bouchard, C., Yamanaka, T., Aoussat, A. (eds.) Proc. of the Kansei Engineeringand Emotion Research Int. Conf. 2010 - KEER 2010. pp. 1–10 (2010)

8. Campos, A., Dignum, F., Dignum, V., Signoretti, A., Magaly, A., Fialho, S.: Aprocess-oriented approach to model agent personality. In: Sierra, C., Castelfranchi,C., Decker, K.S., Sichman, J.S. (eds.) Proc. of the 8th Int. Conf. on AutonomousAgents and Multiagent Systems (AAMAS 2009). pp. 1141–1142. IFAAMAS (2009)

9. Canuto, A.M., Campos, A.M., M.Santos, A., Moura, E.C., Santos, E.B., Soares,R.G., Dantas, K.A.: Simulating working environments through the use ofpersonality-based agents. In: Sichman, J.S., Coelho, H., Rezende, S.O. (eds.) Ad-vances in Artificial Intelligence - IBERAMIA-SBIA 2006, pp. 108–117. No. 4140in LNAI, Springer (2006)

10. Castelfranchi, C., de Rosis, F., Falcone, R., Pizzutilo, S.: Personality traits andsocial attitudes in multi-agent cooperation. Applied Artificial Intelligence 12(7-8),649–675 (1998), special Issue on ‘Socially Intelligent Agents’

11. Connor-Smith, J.K., Flachsbart, C.: Relations between personality and coping:A meta-analysis. Journal of Personality and Social Psychology 93(6), 1080–1107(2007)

12. Dastani, M., Lorini, E.: A logic of emotions: from appraisal to coping. In: Conitzer,Winikoff, Padgham, van der Hock (eds.) Proc. of the 11th Int. Conf. on AutonomousAgents and Multiagent Systems. IFAAMAS (2012)

13. Dryer, C.: Getting personal with computers: How to design personalities for agents.Applied Artificial Intelligence 13(3), 273–295 (1999)

14. Du, H., Huhns, M.N.: Determining the effect of personality types on human-agentinteractions. In: Web Intelligence (WI) and Intelligent Agent Technologies (IAT),2013 IEEE/WIC/ACM Int. Joint Conf.s on. vol. 2, pp. 239–244. IEEE (2013)

15. Durupinar, F., Allbeck, J., Pelechano, N., Badler, N.: Creating crowd variation withthe OCEAN personality model. In: Padgham, Parkes, Muller, Parsons (eds.) Proc.of the 7th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS2008), pp. 1217–1220. IFAAMAS (2008)

16. Egges, A., Kshirsagar, S., Magnenat-Thalmann, N.: Generic personality and emo-tion simulation for conversational agents. Computer Animation and Virual Worlds15, 1–13 (2004)

17. Furnham, A.: The big five versus the big four: The relationship between the myers-briggs type indicator (mbti) and neo-pi five factor model of personality. Personalityand Individual Differences 21(2), 303–307 (1996)

18. Jed, E.M., Pallamin, N., Dugdale, J., Pavard, B.: Modelling character emotion inan interactive virtual environment. In: Proc. of AISB, Leeds, UK (2004)

19. Jones, H., Saunier, J., Lourdeaux, D.: Personality, emotions and physiology in a bdiagent architecture: The pep ->> bdi model. In: Pro. of the 2009 IEEE/WIC/ACMInt. Conf. on Intelligent Agent Technology, IAT 2009. pp. 263–266. IEEE (2009)

20. Lutzenberger, M., Albayrak, S.: Current frontiers in reproducing human driverbehavior. In: Proc. of the 46th Summer Computer Simulation Conf. 2014, pp. 514–521 (2014)

21. Masuch, M., Hartman, K., Schuster, G.: Emotional agents for interactive environ-ments. In: Creating, Connecting and Collaborating through Computing, 2006. C5’06. The Fourth Int. Conf. on. pp. 96–102 (Jan 2006)

22. McCrea, R.R., John, O.P.: An introduction to the five-factor model and its appli-cations. Journal of Personality 60(2), 175–215 (1992)

23. Myers, I.B., Byers, P.B.: Gifts Differing: Understanding Personality Type. NicholasBrealey Publishing, 2 edn. (1995)

24. Neto, A.F.B., da Silva, F.S.C.: On the construction of synthetic characters withpersonality and emotion. In: da Rocha Costa, A.C., Vicari, R., Tonidandel, F.(eds.) Advances in Artificial Intelligence SBIA 2010, Lecture Notes in ComputerScience, vol. 6404, pp. 102–111. Springer (2010)

25. Ozer, D.J., Benet-Martınez, V.: Personality and the prediction of consequentialoutcomes. Annu. Rev. Psychol. 57, 401–421 (2006)

26. Pereira, D., Oliveria, E., Moreira, N.: Formal modeling of emotions in BDI agents.In: Sadri, F., Satoh, K. (eds.) Computational Logic in Multi-Agent Systems, Lec-ture Notes in Computer Science, vol. 5056, pp. 62–81. Springer (2008)

27. von der Putten, A.M., Kramer, N.C., Gratch, J.: How our personality shapes ourinteractions with virtual characters — implications for research and development.In: Allbeck, J., Badler, N., Bickmore, T., Pelachaud, C., Safonova, A. (eds.) Intel-ligent Virtual Agents, Lecture Notes in Computer Science, vol. 6356, pp. 208–221.Springer (2010)

28. Rao, A.S., Georgeff, M.P.: BDI agents: From theory to practice. In: Lesser, V.,Gasser, L. (eds.) Proc. of the First Int. Conf. on Mutiagent Systems (ICMAS1995). pp. 312–319. AAAI, The MIT Press (1995)

29. Salvit, J., Sklar, E.: Toward a myers-briggs type indicator model of agent behaviorin multiagent teams. In: Bosse, T., Geller, A., Jonker, C.M. (eds.) Multi-Agent-Based-Simulation XI, Int. Workshop, MABS 2010, Revised Selected Papers. pp.28–43. No. 6532 in Lecture Notes in Artificial Intelligence, Springer, (2011)

30. Salvit, J., Sklar, E.: Modulating agent behavior using human personality type. In:Proc. of the Workshop on Human-Agent Interaction Design and Models (HAIDM)at Autonomous Agents and MultiAgent Systems (AAMAS). pp. 145–160 (2012)

31. Steunebrink, B.: The Logical Structure of Emotions. Ph.D. thesis, Utrecht Univer-sity (2010)

32. Talman, S., Hadad, M., Gal, Y., Kraus, S.: Adapting to agents’ personalities innegotiation. In: Pechoucek, M., Steiner, D., Thompson, S. (eds.) Proc. of the 4th

Int. Joint Conf. on Autonomous Agents and Multiagent Systems (AAMAS). pp.383–389. ACM (2005)

33. Wilks, L.: The stability of personality over time as a function of personality traitdominance. Griffith University Undergraduated Student Psychology Journal 1, 1–9(2009)

34. Wooldridge, M.: Reasoning about Rational Agents. Intelligent Robotics and Au-tonomous Agents, The MIT Press (July 2000)