Modelling of Personality in Agents: From Psychology to Implementation

Page 1

Modelling of Personality in Agents: From Psychology to Implementation Sebastian Ahrndt, Johannes F¨ahndrich, and Sahin Albayrak DAI-Laboratory, Technische Universit¨ at Berlin, Faculty of Electrical Engineering and Computer Science, Ernst-Reuter-Platz 7, 10587 Berlin, Germany sebastian.ahrndt@dai-labor.de (Corresponding author)

Abstract. There is increasing interest in the agent community to integrate the concept of emotions and artificial agents. The spectrum of available solutions reaches from applications and models of emotions to complete axiomatised logics. Despite the rich offer of solutions, available works neglect individual personality as a significant factor for the outcome of emotional behaviour pattern. However, different personalities affect all relevant phases of human decision-making processes. Hence, this paper introduces and discusses existing personality theories and highlights the fact that one of them is widely accepted in psychology and should be adopted by the agent-community. We integrate the characteristics of this personality theory into the life-cycle of BDI agents and discuss two different versions of the BDI algorithm – a naive one and one that balances the commitment between means and ends. The outlined algorithm is implemented as a prototype model in AntMe!, an agentbased simulation environment for behavioural studies. The experiments performed in this environment show that personality indeed affects all relevant phases of the decision-making process, laying the foundations for future empirical studies.

1

Introduction

Over the last years, the agent community presented several approaches to bring emotions to artificial agents. Available solutions reach from modelling and applying 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 decisionmaking in a use-case independent and principle manner. Comparing this rich offer of works with the available literature on agents incorporating another important aspect of human behaviour reveals an interesting gap: The missing link between personality and emotions in agent-based systems. But is it not that our personality affects our emotions and determines our entire behaviour? Following Ozer and Benet-Mart´ınez [25], personality, in fact, is a significant factor for human behaviour and determines the individual outcome of 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 concept of personality itself. In psychology different theories exist that explain the behaviour of humans describing the humans personality along personality traits or types. These theories have in common that each trait/type is a characteristic feature of a human, which can be used to explain the humans behaviour and its motives along patterns of behaviour. Nowadays two big theories about human personality exist [17]: The Five-Factor Model of personality [22] (FFM) and the Myers-Briggs Type Indicator [23] (MBTI). Both theories and the differences between them are discussed in detail in a prior work [2] from the psychological perspective. The argumentation comprises the origin (empirical vs. theoretical), the scale types (traits with continuous scale vs. types representing clusters), the completeness (both incomplete), and the reliability or consistency of assessments according to both theories. Balancing the arguments, we concluded that psychologists 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 characterising an individual. These are openness to experience, which is related to a person’s preference to act inventive, emotional and curious vs. acting consistent, conservative and cautious; conscientiousness, which is related to a person’s preference to act efficient, planned and organised vs. acting easy-going, spontaneous and careless; extraversion, which is related to a person’s preference to act outgoing, action-oriented and energetic vs. acting solitary, inward and reserved; agreeableness, which is related to a person’s preference to act friendly, cooperative and compassionate vs. acting analytical, antagonistic and detached; and neuroticism which is related to a person’s preference to act sensitive, pessimistic and nervous vs. acting secure, emotionally stable and confident. 1.2

Agents with Personality

In research on agent-based systems, models of human personality are comprehensively used for the implementation of (microscopic) traffic simulation frameworks [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 different personalities into single agents influences the behaviour of a crowd. For this simulation the authors applied FFM. Other areas include human-machine interaction [13], in particular conversational agents/virtual humans [4, 16] and life-like characters [5]. The latter outlines three projects that apply two dimensions of FFM (extraversion and aggreeableness). The effects are interpreted in a rulebased or scripted manner. Another branch of research focuses on modelling and examining the effects of personalities on interactions between agents and their environments. In particular, the effects of personalities in cooperative settings. Talman et al. [32] present a work that illustrates the use of a rather simple abstraction of personality types. Personalities of agents are determined through


the two dimensions cooperation and reliability, which are used to measure the helpfulness of an agent. The agents have to negotiate and cooperate as cooperation is an inherent part of the game they play. During repeatedly played games the agents reason about each others helpfulness along the two dimensions. As an effect they try to respond more effectively by customising their behaviour appropriately for different personalities. Campos et al. [8] present a work employing MBTI, which is here restricted to two of its dichotomies. It is integrated into the reasoning process of a BDI agent and the work proves that different personality characteristics lead to varieties in the decision-making process in a simulation specifically designed for the paper’s use-case. In an early work, Castelfranchi et al. [10] present a framework to investigate the effects of personalities on social interactions between agents, such as delegation and help. The agents apply opponent modelling in terms of personality traits to motivate interactions. However, the work discusses personality traits as an abstract concept without relation to psychological theories. The work that is most closely related to our work is presented by Salvit and Sklar [29, 30]. That is the case, because the authors setup an experiment validating 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 the behaviour of each MBTI type is analysed in a simulation environment called the ‘Termite World’. The results underline the hypothesis of the paper that the different personality types act in quite different ways. Discussion There are several works applying personality to different domains such as traffic simulation or the study of cooperative decision making. They frequently 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 and agents with personality there are even some works that slightly touch the topic of personality when working on emotional agents (cf. [7, 19, 24]). Thus approaches are discussing architectural considerations from the software engineering perspective. However, the literature overview also shows that the majority of works uses simplified models of personality that are not based on psychology findings or apply specific traits. Stunningly, reasons for using either the FFM or MBTI are missing in all considered works. 1.3

Motivation and Problem

To conclude, the questions of how to incorporate personality into agents was not satisfiable approached by agent researchers, yet. In fact, influences of personality 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 at the 2nd HAIDM in 2012.


of Castelfranchi et al. [10] as an abstract concept without a relation to psychological findings. The work of Salvit and Sklar [29, 30] discussed and evaluated these influences with respect to the MBTI and concluded ‘that some agent personality types are better suited to particular tasks—the same observation that psychologists make about humans’ [30, p. 147]. As psychologist tend to accept the FFM as personality framework and at the same time tend to refuse the MBTI, the motivation for this work is to confirm the findings of Salvit and Sklar with respect to the FFM. This will finally provide the foundation for the logical formalisation of personality influences on an agents decision making process as claimed as prerequisite to close the gap between emotional agents and agents with personality elsewhere [3]. Remainder In the following, we present an extension of a selected BDI algorithm as well as particulars about the implementation of this algorithm in a multi-agent simulation in Section 2. In doing so, we assess the level to which personality affects the different stages of the sense-plan-act lifecycle. We establish this assessment by means of simulation results in Section 3. Collected simulation results show that the quality by which problems are solved indeed varies with the problem-solvers personality, that is, problem solving can be altered (and somewhat improved) by a careful personality-specific task assignment. We conclude our findings in Section 4.

2

Modelling Agents with the FFM

In the following, we will discuss how the FFM can be integrated into the BDI model of agency [28], a popular model for the conceptualisation of human behaviour. BDI agents separate the current execution of a plan from the activity of 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 the Belief Revision, the Option Generation, the Filter Process, and the Actuation. In our model, the phases of the BDI cycle are influenced by the characteristics of a personality in different ways. For instance, the trait conscientiousness strongly influences the goal-driven behaviour of an agent, whereas the trait extraversion influences the agent’s preference to interact with others. Table 1 lists the influences of the different characteristics of FFM on the different phases of the BDI life-cycle. These influences address the intensity by which one personality trait influences a phase and thus (only) highlights the traits that are most influential. Indeed, this classification is discussable as it reflects our own interpretation of the FFM traits in comparison with the BDI phases. To substantiate our interpretation we took in account works of different authors investigating the relation between personalities and behaviour types (e.g., [14, 27]). Furthermore, we learned about the influences by experiments that provide findings for the 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 Belief Revision Option Generation Filter Process Actuation

2.1

C

E

× × ×

× × ×

× ×

A

N

× × ×

× ×

Personality and the naive BDI lifecycle

To explain the model, we will use the following syntax introduced by Michael Wooldridge [34, pp. 69–90] for the ‘Logic Of Rational Agents’ (LORA): – P : P er is the collection of personality traits of the agent; – ρ : P ercepts is the information that the agent perceives/receives in its environment; – 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 influence during the different stages. All personality traits are considered during the process. Furthermore, we assume that the personality does not change during the life-cycle of an agent. That is based on the finding that we as humans have a stable personality over our lifespan as adults [33]. The cycle starts with the perception of information. During this stage the agent receives new information from the environment (Env) using its sensors, which also comprises messages (M sg) from other agents (communication acts). The perception is not affected by the personality, as humans are not able to restrict their perception during the cognition. This is a deliberate process taking place in the next step of the cycle. Formally, the signature of the perception function percept is defined as: percept : Env × M sg → P ercepts. The next step of the BDI life-cycle is the Belief Revision. That means that given perceptions (ρ) are computed with respect to the current personality (P ) to update the actual beliefs (B). The belief revision function beliefRevision is defined as: belief Revision : ℘(Bel) × P ercepts × P er → ℘(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 making process. Input: Binit , Iinit , P ; Output: 1: B ← Binit , I ← I init 2: while true do 3: ρ ← percept(Env, M sg) 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) do 9: α ← hd(π) 10: execute(α, P ) 11: π ← tail(π) 12: end while 13: end while

and facts that were received via communication. In our model the O and A characteristics influence this phase most frequently, as they influence the interpretation what the new measurement means for the agent and how trustful the agent is when receiving information from others. One essential reason to distinguish between perceptions/beliefs derived from the environment and perceptions/beliefs derived from other agents is the characteristic of the personality trait agreeableness, which indicates the preference to trust others.2 We implemented this behaviour (the influence of the trait A during the belief revision) for our simulation environment using the characteristic of the personality trait as likelihood. For example, an agent with A = 1.0 always trust information received via 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 and N 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 external influences. This deliberation process is represented by the function options with the following signature: options : ℘(Bel) × ℘(Int) × P er → ℘(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 produce all options available to the agent the influence of the personality is restricted to the persistence of already selected intentions. Again, we implemented the 2

In fact, it might be hard to clearly distinguish the information sources. That is because other agents are part of the environment and the observation of the behaviour of other agents might thus be both an observation of the environment and an (implicit) communication act.


influence by interpreting the traits as likelihood, e.g. an agent with C = 1.0 always maintain an intention as option regardless of the current beliefs about the world. The third stage is the Filter Process where the agent chooses between competing desires and commits to achieve some of them next. The filter process is influenced by the preferences to vary activities over keeping a strict routine and the level of self-discipline (O, C), the need to act in harmony with other agents (A, N) and even the tendency to generally interact with others (E). For example, variations of C influence an agent’s preference to detach the prior selected intentions. As another example, variations of A and E influence an agent’s preference to commit to selfish/altruistic goals. The filter function is defined as: f ilter : ℘(Bel) × ℘(Des) × ℘(Int) × P er → ℘(Int). The personality helps to prioritise the different intentions and for example indicates to what extent an agent acts goal-driven, prefers interaction, varies the activities. It selects the best option from the point of view of the agent based on the current beliefs, with respect to the prior selected option. Again interpreting the traits as likelihood, the filter process is implemented by, e.g., prioritising intentions 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 mainly influenced by the creativity level of the agent (O), the tendency to apply actions in a decent manner (C) and the preference to interact with others (E). The actual plan is then generated for the selected intentions and executed, which is defined as: plan : ℘(Bel) × ℘(Int) × P er → Act∗ . The execution of actions as plan-elements directly influences the environment and the personality indicates how accurate an agent behaves (C). This is a rather vague argument for agents. To set an example, imagine a robot that should perform a motion from one point to another in a specific time frame. The level of conscientiousness then can be used to implement a noise level added to the target location or time frame boarders. Indeed, this seems to be curious when considering artificial agents but is one important difference between humans. The actuation function execute is formally defined as: execute : Act × P er 2.2

Balancing commitments to means and ends

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


this work as: the domain is tick-based, the plans a rather short, and plans executed for an intention are fixed, making the time required to generate a plan negligible. However, using the provided explanation the algorithm can be adapted to produce reactive and single- or open-minded behaviour, which might be either bold or cautious. Algorithm 2 shows one variant of a BDI life-cycle that is not overcommitted 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 making process that is not overcommitted to the means or ends. Input: Binit , Iinit , P ; Output: 1: B ← Binit , I ← I init 2: while true do 3: ρ ← percept(Env, M sg) 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)) do 9: α ← hd(π) 10: execute(α, P ) 11: π ← tail(π) 12: ρ ← percept(Env, M sg) 13: B ← beliefRevision(B, ρ, P ) 14: if reconsider(I, B) then 15: D ← options(B, I, P ) 16: I ← filter(B, D, I, P ) 17: end if 18: if not sound(π, I, B) then 19: π ← plan(B, I, P ) 20: end if 21: end while 22: end while

perception and belief revision stage. This is done as each action takes some execution time and thus the environment might change to a state where the current intention or the current selected plan is not of relevance anymore. The process outlined in Algorithm 1 can not recognise these facts as it strictly executes a once selected plan. Also introduced are some methods that help the agent to decide whether it must reconsider its current intentions or whether the currently selected plan is sound. The inner while condition is further extended with two conditions, which validate whether the intentions were successfully achieved or became impossible.


Taking the above argumentation about the influence of the personalty into account 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 N indicating the emotional stability could influence the impossible check, in terms of ‘more rounds, more stressful’. However, the extension mechanism proposed affects the existing stages of the BDI cycle and we argue that these effects take place in these stages. Making the influence in the condition checks redundant. For example, whether an intention was successfully achieved is recognised in the belief revision and thus will make the effects of C implicit available in the condition check.

3

Experimental Setup and Results

To evaluate the model we implemented it for the multi-agent simulation environment 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 as other 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 environment, updating its beliefs, desires and intentions and executing. The ants are able to sense their location and to recognise whether or not they are transporting food, the location of food, other ants, scent-marks, and enemy within their range of sight. The scent-marks are used to determine what other ants of the own colony are targeting and to highlight the occurrence of enemies. The possible actions are goStraight, goAwayFromPOI, goToPOI, goToNest (‘move actions’), turnToPOI, turnByAngle, turnAround, turnToGoal (‘turn actions’), pick-up and drop-off food, attack, and put scent-mark. Fig. 1a shows a screenshot of the simulation environment. Using the introduced model we expect that the ants’ behaviours vary when adjusting the personality traits. In particular we expect that an ant population with high values in the trait openness (O+) does more exploration more than a population with low values (O-).4 That means that O+ ants are expected to find sugar and apples earlier. At the same time, we expect the O- ants to harvest sugar faster as a consistent behaviour is favourable for this task, which includes walking the same route multiple times.5 We expect that high values in the trait conscientiousness (C+) lead to more collected food, as such ants will not drop food when facing other goals such as attacking/running away from bugs. At the same time, we expect low valued ants (C-) to have a lower chance to starve during the search for food as collecting food is the most important desire. Extroverted ants (E+) are expected to communicate more frequently with other ants by putting scent-marks as markers for the occurrence of sugar, 3

4 5

For further information about the simulation environment the interested reader is referred to http://www.antme.net/. The −, + label represent a value in the interval [0.0, 0.5), [0.5, 1.0] respectively. In other words, openness indicates the choice between exploration and exploitation.


apples and bugs more frequently. However, this effect correlates with the effect of the trait agreeableness, indicating whether an ant trusts information received from other ants (A+) or not (A-). We expect that high valued ants in both traits collect food more frequently. The neuroticism trait indicates the ants’ emotional stability. We expect high valued ants (N+) to avoid dangerous situations such as bugs and hostile ants – resulting in lower numbers of eaten ants and killed bugs. 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 personality traits and the measurable features of an AntMe! simulation. For this we simulated the permutation of the minimum and maximum values for each trait, resulting in 25 = 32 ant populations. The features comprise the collected apples and the collected sugar, the number of eaten and starved ants, and the number of killed bugs. For each permutation the values were averaged over 50 simulation runs, where each simulation run started with the same point of origin of the ant hill, apples, and sugar. Occurrence of bugs is randomised and each deceased ant 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- ants and that this behaviour is most notable for the collected sugar. Still, we postulated 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 populations start approximately 2% earlier with the collection, but collect food slower than O- ant populations. The lower part of Table 2 lists the results of all ant populations representing a permutation of the minimal and maximum values of the personality traits. It is emphasised that different types of personality lead to different simulation results. For example, an ant population with maximum values (1,1,1,1,1) collects more apples and sugar, kills fewer bugs and loses fewer ants through bugs than an ant population with minimum values (0,0,0,0,0). Still, for the latter a lower number of starved ants can be observed. Here, the traits E and A influence the occurrence 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 new percepts as collecting food is the most important goal for the ants. The trait N affects the fight behaviour of the ants leading to fewer/more eaten ants/killed bugs, respectively. The effects of the personality traits are also visible in the paths an ant population 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 acting 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 round artefacts highlighting the occurrence of apples – collecting apples is a cooperative task.


Table 2: Correlation matrix between measured items and personality traits (upper part) and collected information for a set of ant populations. Lowest and highest measured items ar highlighted in bold. O C E A N

Apple

Sugar

Eaten

Starved

Bugs

-0.068 0.545 -0.150 0.261 0.305

-0.444 0.425 0.072 0.501 0.114

-0.043 -0.454 0.002 -0.430 -0.436

-0.209 0.893 -0.119 0.107 0.125

0.027 -0.027 -0.009 -0.554 -0.554

values below are ordered according to the OCEAN acronym (0,0,0,0,0) (0,0,0,0,1) (0,0,0,1,0) (0,0,0,1,1) (0,0,1,0,0) (0,0,1,0,1) (0,0,1,1,0) (0,0,1,1,1) (0,1,0,0,0) (0,1,0,0,1) (0,1,0,1,0) (0,1,0,1,1) (0,1,1,0,0) (0,1,1,0,1) (0,1,1,1,0) (0,1,1,1,1) (1,0,0,0,0) (1,0,0,0,1) (1,0,0,1,0) (1,0,0,1,1) (1,0,1,0,0) (1,0,1,0,1) (1,0,1,1,0) (1,0,1,1,1) (1,1,0,0,0) (1,1,0,0,1) (1,1,0,1,0) (1,1,0,1,1) (1,1,1,0,0) (1,1,1,0,1) (1,1,1,1,0) (1,1,1,1,1) ( 12 , 12 , 12 , 12 , 12 )

8.4 19.5 19.8 19.4 8.2 18.6 16.7 16.1 19.0 19.4 19.7 19.3 19.0 19.3 16.5 16.0 8.5 16.2 16.8 16.2 7.9 15.7 16.2 15.8 19.2 19.5 19.6 19.4 18.8 19.1 19.4 19.3 9.7

18.4 81.0 83.2 78.2 10.6 43.1 113.1 108.3 75.6 88.3 90.1 86.8 52.9 58.0 181.0 175.7 8.1 46.3 48.1 44.5 6.5 28.9 40.1 39.9 70.5 65.0 69.6 66.7 47.2 50.4 78.6 75.8 17.1

281.6 84.7 82.8 84.7 284.8 97.2 83.0 83.4 117.4 54.5 54.3 55.7 98.3 51.5 65.9 64.2 285.4 77.1 82.7 80.1 283.9 74.0 74.4 75.0 90.7 54.6 54.3 52.1 99.7 52.6 55.5 54.2 270.8

6.0 130.9 131.5 131.6 3.5 39.8 55.1 55.3 146.9 204.8 203.5 203.1 162.9 204.7 174.6 175.9 0.0 0.0 0.0 0.0 0.1 0.0 0.0 0.0 167.6 193.5 193.4 197.1 153.0 193.3 184.5 188.4 19.5

2.5 0.0 0.0 0.0 1.8 0.0 0.0 0.0 3.3 0.0 0.0 0.0 2.1 0.0 0.0 0.0 3.0 0.0 0.0 0.0 3.5 0.0 0.0 0.0 1.6 0.0 0.0 0.0 2.7 0.0 0.0 0.0 1.5


(a) Map

(b) (0,0,0,0,0)

(c) ( 12 , 12 , 21 , 12 , 12 )

(d) (1,1,1,1,1)

(e) (0,1,1,1,0): Max. (f) (1,0,1,0,0): Min. (g) (0,0,0,1,0): Min. (h) (0,1,0,0,1): Max. sugar sugar starved starved

Fig. 1: Fig. 1a shows the map used for the single population simulations. The occurrence of food (apples in green, sugar in white) and the location of the ant hill are fixed. The ants goal is to collect as much food as possible and not to die either by starving or by fighting against bugs (blue). Fig. 1b–1h show the cumulated paths of seven ant populations. As the map is fixed a comparable structure originates. Still, the effects of exploration vs. exploitation are visible (covered area, curious behaviour, broader paths). The artefacts denote the visibility range of the ants and the points apples are spawned, giving an indication of the effects of 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 added to the BDI lifecycle can be interpreted as personality traits and the resulting behavioural change of the agents can be interpreted as personality. In addition, we have shown that different personality traits affect the result of the simulation and 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 personality traits available through the Five-Factor Model. We have also learned that such parameters can influence the behaviour of agents in a domain independent manner and that one challenge is the task-dependent interpretation of the effect of a personality. Finally, the experiment confirmed the finding of Salvit and Sklar [30] that the interpretation of the parameters as personality traits results in (personality-)consistent behaviour of agents with respect to the Five-Factor Model instead of the MBTI. The implication is that the task performance of problem-solvers can be improved by carefully assigning personality-specific tasks. To show that, we can use the derived results to determine which population performs more accurate for a


specific objective. In Table 2 the minimal and maximal values that were reached by the populations for the different measured categories are highlighted. Is the objective to collect as much sugar as possible the population (0,1,1,1,0) would be 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 that collect 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 attack bugs frequently leading to a high amount of eaten ants. Another example is the objective to let as least ants starve as possible (ants starve if they not rest; rest means staying in the anthill). The paths of one of the populations that are not starving is shown in Fig. 1g. Here the concentration on staying in near distance to the anthill becomes visible. In contrast, Fig. 1h shows the path heatmap of an ant population where the individuals avoid periods of rest starving when exploring the map.

3.2

First Empirical Results

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

Table 3: Ten real personalities with corresponding simulation results. OCEAN (0.85, (0.49, (0.63, (0.81, (0.59, (0.59, (0.95, (0.70, (0.95, (0.71, 6

0.70, 0.95, 0.63, 0.84, 0.59, 0.69, 0.95, 0.35, 0.69, 0.71,

0.47, 0.31, 0.65, 0.73, 0.44, 0.24, 0.49, 0.70, 0.76, 0.51,

0.34, 0.70, 0.71, 0.60, 0.64, 0.43, 0.83, 0.66, 0.90, 0.64,

0.83) 0.48) 0.64) 0.46) 0.60) 0.49) 0.73) 0.56) 0.83) 0.41)

Apple

Sugar

Eaten

Starved

Bugs

13.4 17.4 13.10 13.42 11.64 10.24 18.96 10.36 18.20 11.50

25.8 49.9 26.68 27.34 23.14 19.26 60.66 17.88 51.32 21.88

233.6 166.1 239.18 234.26 255.48 266.32 115.66 265.76 138.18 256.84

47.1 104.5 42.58 46.14 29.60 21.64 144.90 21.26 123.92 28.82

0.7 1.3 0.70 0.90 1.24 1.40 0.26 1.46 0.02 0.84

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


The intention to use realistic personalities is to compare these simulation results with the expectation of the participants in an uninformed and informed stage. To do so we explained the simulation environment and the measurable items to the participants and asked them to formulate their expectations before they were informed about their personalities and afterwards. Until now, the analysis 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 works that integrate personality as a factor for agent-based behaviour. We showed that there is a gap between the progress that was made for emotional agents and that there is a missing link between both (essential) human behavioural processes. Based on this finding, we took the Five-Factor Model of personality into account and discussed how it can be integrated into the BDI reasoning process. We demonstrated the applicability of our approach by means of AntMe!, an agentbased simulation framework, which provides a completely adaptable test-bed for behavioural studies. Despite the fact that we simulated ants, we were able to show that personality affects all relevant phases of agent decision-making processes and conclude that personality-specific task assignment can alter and/or improve the quality in which problems are being solved. Having done that, we were able to confirm the findings of Salvit and Sklar [29, 30] with respect to the FFM, which is the initial intention to present this work. It is important to mention, that we presented a stepping-stone rather than a holistic solution. A more comprehensive implementation should address several issues, most importantly: The impact of one particular personality trait is always subject 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 might act rude, when writing emails or chatting to people they never met. Finding concepts for this particular characteristic of human behaviour and integrating these concepts with existing emotional agent approaches is an open topic and requires both, theoretical work and user-studies. In another work [3], we briefly introduce the first steps towards a theoretical integration by presenting thoughts on a logical formalisation of the here presented algorithm. The basic idea is to integrate personality as an own modal connectivity in the BDI life-cycle.

References 1. Adam, C.: Emotions: from psychological theories to lologic formalization and Implementation in a BDI agent. Ph.D. thesis, L’ Institute National Polytechnique de Toulouse (2007) 2. Ahrndt, S., Aria, A., F¨ ahndrich, J., Albayrak, S.: Ants in the OCEAN: Modulating agents with personality for planning with humans. In: Bulling, N. (ed.) Multi-Agent Systems (EUMAS 2014), pp. 1–16. Lecture Notes in Artificial Intelligence, Springer (December 2014)


3. Ahrndt, S., Faehndrich, J., L¨ utzenberger, M., Albayrak, S.: Modelling of personality in agents: From psychology to logical formalisation and implementation. In: Bordini, Elkind, Weiss, Yolum (eds.) Proc. of the 14th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2015), pp. 1–2. IFAAMAS (May 2015) 4. Allbeck, J., Badler, N.: Toward representing agent behavior modified by personality and emotion. In: Proc. of the Workshop on Embodied Conversational Agents at the 1st Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS). ACM (2002) 5. Andre, E., Klesen, M., Gebhard, P., Allen, S., Rist, T.: Integrating models of personality and emotions into lifelike characters. In: Paiva, A., Martinho, C. (eds.) Proc. of the Workshop on Affect in Interactions Towards a new Generation of Interfaces. pp. 136–149 (2000) 6. Baumert, A., Schmitt, M.: Personality and information processing. European Journal of Personality 26(2), 87–89 (2012) 7. Bevacqua, E., de Sevin, E., Pelachaud, C., McRorie, M., Sneddon, I.: Building credible agents: Behaviour influenced by personality and emotional traits. In: L`evy, P., Bouchard, C., Yamanaka, T., Aoussat, A. (eds.) Proc. of the Kansei Engineering and Emotion Research Int. Conf. 2010 - KEER 2010. pp. 1–10 (2010) 8. Campos, A., Dignum, F., Dignum, V., Signoretti, A., Magaly, A., Fialho, S.: A process-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 Autonomous Agents 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 of personality-based agents. In: Sichman, J.S., Coelho, H., Rezende, S.O. (eds.) Advances in Artificial Intelligence - IBERAMIA-SBIA 2006, pp. 108–117. No. 4140 in LNAI, Springer (2006) 10. Castelfranchi, C., de Rosis, F., Falcone, R., Pizzutilo, S.: Personality traits and social 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 Autonomous Agents 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-agent interactions. 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 with the OCEAN personality model. In: Padgham, Parkes, M¨ uller, Parsons (eds.) Proc. of the 7th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2008), pp. 1217–1220. IFAAMAS (2008) 16. Egges, A., Kshirsagar, S., Magnenat-Thalmann, N.: Generic personality and emotion simulation for conversational agents. Computer Animation and Virual Worlds 15, 1–13 (2004) 17. Furnham, A.: The big five versus the big four: The relationship between the myersbriggs type indicator (mbti) and neo-pi five factor model of personality. Personality and Individual Differences 21(2), 303–307 (1996)


18. Jed, E.M., Pallamin, N., Dugdale, J., Pavard, B.: Modelling character emotion in an interactive virtual environment. In: Proc. of AISB, Leeds, UK (2004) 19. Jones, H., Saunier, J., Lourdeaux, D.: Personality, emotions and physiology in a bdi agent architecture: The pep ->> bdi model. In: Pro. of the 2009 IEEE/WIC/ACM Int. Conf. on Intelligent Agent Technology, IAT 2009. pp. 263–266. IEEE (2009) 20. L¨ utzenberger, M., Albayrak, S.: Current frontiers in reproducing human driver behavior. 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 environments. 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 applications. Journal of Personality 60(2), 175–215 (1992) 23. Myers, I.B., Byers, P.B.: Gifts Differing: Understanding Personality Type. Nicholas Brealey Publishing, 2 edn. (1995) 24. Neto, A.F.B., da Silva, F.S.C.: On the construction of synthetic characters with personality and emotion. In: da Rocha Costa, A.C., Vicari, R., Tonidandel, F. (eds.) Advances in Artificial Intelligence SBIA 2010, Lecture Notes in Computer Science, vol. 6404, pp. 102–111. Springer (2010) 25. Ozer, D.J., Benet-Mart´ınez, V.: Personality and the prediction of consequential outcomes. 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, Lecture Notes in Computer Science, vol. 5056, pp. 62–81. Springer (2008) 27. von der P¨ utten, A.M., Kr¨ amer, N.C., Gratch, J.: How our personality shapes our interactions with virtual characters — implications for research and development. In: Allbeck, J., Badler, N., Bickmore, T., Pelachaud, C., Safonova, A. (eds.) Intelligent 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 (ICMAS 1995). pp. 312–319. AAAI, The MIT Press (1995) 29. Salvit, J., Sklar, E.: Toward a myers-briggs type indicator model of agent behavior in multiagent teams. In: Bosse, T., Geller, A., Jonker, C.M. (eds.) Multi-AgentBased-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 University (2010) 32. Talman, S., Hadad, M., Gal, Y., Kraus, S.: Adapting to agents’ personalities in negotiation. 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 trait dominance. Griffith University Undergraduated Student Psychology Journal 1, 1–9 (2009) 34. Wooldridge, M.: Reasoning about Rational Agents. Intelligent Robotics and Autonomous Agents, The MIT Press (July 2000)


Turn static files into dynamic content formats.

Create a flipbook
Issuu converts static files into: digital portfolios, online yearbooks, online catalogs, digital photo albums and more. Sign up and create your flipbook.