universidade tÉcnica de lisboa instituto superior … · conjuntos de adjectivos é a teoria de...

101
UNIVERSIDADE TÉCNICA DE LISBOA INSTITUTO SUPERIOR TÉCNICO Synthetic Agents with Personality Tiago Manuel Salsinha Doce Dissertação para a obtenção do Grau de Mestre em Engenharia Informática e de Computadores Júri Presidente: Nuno Joao Neves Mamede, Departamento de Engenharia Informática (DEI) Orientadora: Ana Maria Severino de Almeida e Paiva, DEI Co-Orientador: Rui Filipe Fernades Prada, DEI Vogal: Maria Luisa Torres Ribeiro Marques da Silva Coheur, DEI Setembro de 2009

Upload: nguyentuyen

Post on 12-Feb-2019

217 views

Category:

Documents


0 download

TRANSCRIPT

UNIVERSIDADE TÉCNICA DE LISBOA

INSTITUTO SUPERIOR TÉCNICO

Synthetic Agents with Personality

Tiago Manuel Salsinha Doce

Dissertação para a obtenção do Grau de Mestre em

Engenharia Informática e de Computadores

Júri

Presidente: Nuno Joao Neves Mamede, Departamento de Engenharia Informática

(DEI)

Orientadora: Ana Maria Severino de Almeida e Paiva, DEI

Co-Orientador: Rui Filipe Fernades Prada, DEI

Vogal: Maria Luisa Torres Ribeiro Marques da Silva Coheur, DEI

Setembro de 2009

ii

Título: Agentes Sintéticos com Personalidade

Nome: Tiago Manuel Salsinha Doce

Mestrado em: Engenharia Informática e de Computadores

Orientador: Professora Doutora Ana Maria Severino de Almeida e Paiva

RESUMO:

PALAVRAS-CHAVE: Credibilidade, Agentes Sintéticos, Personalidade, Emoções, Big Five, Traço,

Abertura à Experiência, Responsabilidade, Extroversão, Amabilidade, Neuroticismo.

Na era da globalização, aspectos como a individualidade e a personificação tornam-se cada vez

mais importantes em sistemas virtuais. Uma forma de o conseguir é endereçando o problema da

personalidade em agentes sintécticos. Para isso, é necessário olhar para como a personalidade

pode ser representada e que aspectos podem e devem ser analisados e introduzidos num agente

inteligente de forma a criar personalidade de uma forma credível. Quanto à estrutura da

personalidade, uma teoria bem aceite que define personalidade como sendo um conjunto de

predisposições para agir de determinadas formas, estável através to tempo, e descrita com

conjuntos de adjectivos é a teoria de traços Big Five. Esta teoria define cinco traços de

personalidade que descrevem a personalidade num todo, Abertura à Experiência (Openness to

Experience), Responsabilidade (Conscientiousness), Extroversão (Extroversion), Amabilidade

(Agreeableness), Neuroticismo (Neuroticism), fornecendo uma forma sistemática the representar

personalidades, tornando-a adequada para a incorporação num agente inteligente. De todos os

aspectos da personalidade que influenciam o comportamento de uma pessoa, ligámos quatro

deles a quatro principais características de um agente: emoções, processo de coping, processo

de planeamento e expressão corporal. Assim, criámos um agente autónomo e emocional, capaz

de raciocinar e agir baseado em objectivos, planos e acções e influenciámos o seu

comportamento com um sistema de personalidade. Estão assim dependentes da personalidade

do agente, as emoções “sentidas”, bem como a estratégia de coping escolhida, o plano de

execução preferido e os gestos e expressões faciais realizados pelo agente. Para testar a

aplicação, o agente criado foi testado num cenário concreto e os resultados mostram que os

utilizadores conseguiram identificar a personalidade do agente em termos gerais, avaliando os

traços do agente como sendo próximos dos traços reais.

iii

Title: Synthetic Agents with Personality

Name: Tiago Manuel Salsinha Doce

Abstract:

KEY-WORDS: Believability, Synthetic Agents, Personality, Emotions, Big Five, Trait, Openness to

Experience, Conscientiousness, Extroversion, Agreeableness, Neuroticism.

In the era of globalization, concepts such as individualization and personalization become more

and more important in virtual systems. One way to accomplish this is to address the issue of

personality in the computational systems created, in particular, in synthetic agents. To do this, one

needs to look at how personality can be represented and what aspects of personality can and

should be analyzed and introduced in an intelligent agent in order to create personality in a

believable manner. Regarding the structure of personality, a well accepted theory (the Big Five

trait theory) defines it as being a set of predispositions to act in certain ways, stable through time,

and described with sets of adjectives.. This theory describes five traits that define the whole

personality, Openness to Experience, Conscientiousness, Extroversion, Agreeableness and

Neuroticism, providing us with a systematic way to represent personality, making it adequate for

incorporation in an intelligent agent. From all the personality aspects that are said to influence a

person's behavior, we link four of them to four main agent characteristics: emotions, coping

process, planning process and bodily expression. So, we created an autonomous emotional

agent, able to reason and act based on goals, plans and actions, and influence the agent's

behavior with a personality system. It will then be dependent on the agent's personality, the

emotions felt, as well as the coping strategy chosen, the preferred plan to meet a specific goal and

the gestures and facial expressions performed by the agent. In order to evaluate the system, the

agent created was tested in a concrete scenario and the results show that users could identify the

agent's personality in general terms, perceiving the agent's traits as being close to the actual

original traits.

iv

Acknowledgements

First of all I would like to thank my supervisor, Professora Ana Paiva, for all the guidance she

provided me, and for leading me when I did not know what to do, as well as Professor Rui Prada

for the usefull tips.

I also thank João Dias for his support and patience, spending time helping me with things he had

no obligation of.

I cannot forget my classmate and friend Samuel Matos for his collaboration and companionship in

the adventure that was developing our theses.

Finally, I would like to thank my family and friends for all the support and for helping me enjoy this

last year.

v

Table of Contents

1 Introduction 1

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1

1.2 The Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.3 Thesis Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Related Work 6

2.1 Background. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1.1 Gordon Allport as the father of personality theory . . . . . . . . . . . . 7

2.1.2 Raymond Cattell and factor analyses . . . . . . . . . . . . . . . . . . .7

2.1.3 Hans Eysenck . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8

2.1.4 The Big Five . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8

2.1.5 Emotions and Personality . . . . . . . . . . . . . . . . . . . . . . . . .11

2.2 Personality in Intelligent Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.2.1 Cybercafé application . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.2.2 An intelligent agent's processes . . . . . . . . . . . . . . . . . . . . . .18

2.2.3 Goals and Plans to represent Personality . . . . . . . . . . . . . . . . .19

2.2.4 Agent Personality Traits in Virtual Environments Based on Appraisal Theory

Predictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.2.5 A multilayer personality model . . . . . . . . .. . . . . . . . . . . . . . 23

2.2.6 Emotion and Personality in Driver Assistance Systems . . . . . . . . . 24

2.3 Work Implications. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

vi

3 An Architecture for Agents with Personality 27

3.1 Personality and emotions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.2 Personality and planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3.3 Personality and coping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.4 Personality and expressivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4 Implementation 47

4.1 FAtiMA and GRETA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.1.1 FAtiMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.1.2 GRETA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4.2 Implementing Personality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

4.2.1 Implementing personality in the appraisal process . . . . . . . . . . . . 53

4.2.2 Implementing personality in the planning process . . . . . . . . . . . . 53

4.2.3 Implementing personality in the coping process . . . . . . . . . . . . . 54

4.2.4 Implementing personality in bodily expressivity . . . . . . . . . . . . . .55

4.3 A practical example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .56

4.3.1 The agent's role . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

4.3.2 Defining goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

4.3.3 Actions and inference moments . . . . . . . . . . . . . . . . . . . . . . 60

4.3.4 Relating actions to expressions . . . . . . . . . . . . . . . . . . . . . . 61

4.3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

4.3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .63

5 Evaluation 64

5.1 Test Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .64

5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .66

5.3 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .70

6 Conclusion 72

References 74

A Questionnaire 80

B Statistical Data 84

vii

List of Figures

2.1 Trait dimensions and their opposite features . . . . . . . . . . . . . . . . . . . . . . . . 10

3.1 Architecture of a Synthetic Agent with Personality . . . . . . . . . . . . . . . . . . . . . 27

4.1 The personality influencing agent's acting process . . . . . . . . . . . . . . . . . . . . . 48

4.2 Agent with a cheerful personality (O-3; C-3; E-5; A-4; N-1) . . . . . . . . . . . . . . . . . 62

4.3 Agent with a neurotic personality (O-3; C-3; E-3; A-3; N-5) . . . . . . . . . . . . . . . . . 62

5.1 Mean values for the four analyzed traits in the melancholic personality. . . . . . . . . . . 67

5.2 Mean values for the four analyzed traits in the sanguine personality . . . . . . . . . . . . 68

5.3 Mean values for the four analyzed traits in the choleric personality. . . . . . . . . . . . . 69

5.4 Mean values for the four analyzed traits in the phlegmatic personality . . . . . . . . . . . 70

B.1 Mean and deviation for Conscientiousness evaluated by users for the melancholic agent . 84

B.2 Mean and deviation for Extroversion evaluated by users for the melancholic agent . . . . 85

B.3 Mean and deviation for Agreeableness evaluated by users for the melancholic agent . . . 85

B.4 Mean and deviation for Neuroticism evaluated by users for the melancholic agent . . . . . 86

B.5 Mean and deviation for Conscientiousness evaluated by users for the sanguine agent . . . 86

B.6 Mean and deviation for Extroversion evaluated by users for the sanguine agent . . . . . 87

B.7 Mean and deviation for Agreeableness evaluated by users for the sanguine agent . . . . 87

B.8 Mean and deviation for Neuroticism evaluated by users for the sanguine agent . . . . . . 88

B.9 Mean and deviation for Conscientiousness evaluated by users for the phlegmatic agent . . 88

B.10 Mean and deviation for Agreeableness evaluated by users for the phlegmatic agent . . . 89

B.11 Mean and deviation for Extroversion evaluated by users for the phlegmatic agent . . . .89

B.12 Mean and deviation for Neuroticism evaluated by users for the phlegmatic agent . . . . 90

B.13 Mean and deviation for Conscientiousness evaluated by users for the choleric agent . . 90

B.14 Mean and deviation for Extroversion evaluated by users for the choleric agent . . . . . 91

B.15 Mean and deviation for Agreeableness evaluated by users for the choleric agent . . . . 91

B.16 Mean and deviation for Neuroticism evaluated by users for the choleric agent . . . . . . 92

List of Tables

Table 3.1 Influence of the Big Five personality traits on the 22 emotions of the OCC theory. . . 35

List of Equations

Equation 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

Equation 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

Equation 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

Equation 4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

Equation 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

Equation 6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

Equation 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

Equation 8 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

Equation 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

Equation 10 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .37

Equation 11 . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

Equation 12 . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

Equation 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .39

Equation 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

Equation 15 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Equation 16 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .45

Equation 17 . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Equation 18 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .45

Equation 19 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Equation 20 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Equation 21 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Equation 22 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .45

Equation 23 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

Equation 24 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

viii

10

Chapter 1

Introduction

1.1 Motivation

In the era of globalization, where the number of individuals and their different characteristics has

increased enormously, more and more aspects such as personalization and individuality become

relevant in human-machine interaction. Every day, people get in touch with a great variety of

unique, personal characteristics of each other, leading to the need to provide them with virtual

entities that can meet those same characteristics. This can help in fulfilling the expectations about

the behaviour of another's, created by a world where every culture and/or individual is important.

Also, with the great amount of information available today, it is easy to create individual

preferences about the way one interacts in certain contexts, and these preferences can be met by

personalizing those interactions accordingly. In addition, these human-machine interactions are

granted with believability, as this feature is determinant in the success of the personalization of the

virtual entity involved. Virtual entities are more believable if they have more realistic, believable

personalities [4], [5]. This concept of believability is, and has been for a long time, pursuit by many

scientists, who wish to create entities that can speak, think and “feel” in a believable way, i. e., in a

way that could provide to the viewers the illusion of life [2]. By making the agents believable, we

can improve the interaction between humans and machines in several areas, including the

entertainment and the educational ones [3]. Mateas [9] states that children are more willing to

learn with the use of believable agents because these agents increase the motivation and the

engagement of the children. Also, it is easy to acknowledge the major and diverse possibilities that

the introduction of believable personalized synthetic characters in the area of games can create.

1

2

With all of this, the importance of a system that can create an assigned personality in a character

and express it becomes clear.

Personality can be defined as a dynamic and organized set of characteristics possessed by a

person that uniquely influences his or her cognitions, motivations, and behaviours in various

situations. [7]

By incorporating personality in synthetic agents, we are trying to give them a set of characteristics

belonging to certain patterns that will result in a coherent behaviour, which in turn increases their

believability [6]. Also, it allows for individuality of such agents which in turn allows for more

believable simulations, and by creating different personalities, agents can be different. Moreover,

this set of personal characteristics is also responsible for improved human-machine interactions,

as humans respond to machines that show personality characteristics as if they had indeed

personality, thus applying the same social rules with machines as the ones they apply when

interacting with other humans [8]. As such, it makes sense to create a consistent model of

personality, that can be as coherent and as believable as possible. This way, the interaction

between humans and machines can also become more familiar [37].

The agents we are trying to create, will, therefore, have personality. However, to do that, their

emotional dispositions and states also have to be represented. These emotional states (and

personalities) can be expressed as our agents interact with each other in scenarios, and they will

be given a certain degree of improvisation in those scenarios. This will also make them more

believable given that improvisation is a social act, where the characters interact in an entertaining

way without being predictable [29].

1.2 The Problem

Given the context just presented, the main questions that will be pursuit in this thesis are:

How can we create synthetic agents with personality? What aspects of the agents should we use

to represent personality?

To answer these questions, it is necessary to know what personality is and how it can be defined,

represented and structured. This can be done by looking at personality theories. There have been

numerous theories regarding personality, including Freud's psychodynamic theory, Jung's

3

psychological types, and trait theories. It is not the concern of this thesis to provide an extensive

review of all the theories in the psychology area, so, we will concentrate our work in one of the

most accepted theories, the trait theories. These theories define personality as a set of

characteristics of the individual that are relatively stable, instead of making the personality

dependent on the context of the situation [60]. According to trait theories, personality can be

described through a wide range of personal features belonging to the individual, such as sympathy

and dominance [10]. This way, it is easier to define an agent's personality, as it concerns the agent

alone.

Given the chosen theory, and in order to put personality in synthetic agents, there are a few

features of the agent we can manipulate:

Emotional state. Personality plays a strong role on our emotional appraisal, having a huge

influence on the emotions we feel, and their intensity [26]. It is known that certain

personalities are associated with certain types of emotions [56]. Therefore, we can define

a relationship between a synthetic agent's personality and the emotions he or she “feels”,

making the agent show emotions according to the attributed personality.

Coping process. Another part of an individual's behaviour is the coping strategies he or

she applies in daily situations. Dealing with problems, individuals react with different

coping strategies, and personality has been shown to influence this process [61], [62],

[63], [64]. We will analyse the different coping strategies and establish a relationship

between an agent's personality and the coping strategies he applies.

Planning. An important part of an agent's mind is its planning. The way the agent

constructs actions to achieve its goals can vary depending on its personality. It is

necessary to look into this feature in detail, so that it can be influenced by the personality

of a synthetic agent.

Body expression. One of the most important parts of personality expression is the way we

gesture and our facial expressions. It is through this expressivity that we make the most

assumptions about one's personality [65]. Therefore it is necessary to study the ways

personality affects the body and facial expressions of an individual and propose a

mapping between the two.

4

The Hypothesis

By using the described above set of features as elements to express personality, users are able to

identify personality differences in synthetic characters.

Work Overview

In order to create synthetic agents with personality we first need an architecture for the agent's

mind. Several types of architecture were studied and used by different researchers. Although other

architectures were analysed, we have chosen one that is based on goals, intentions and plans (the

FAtiMA architecture)

To show the agent's behaviour it is necessary some kind of interface. Among all the possibilities it

made more sense for this work to use an embodied conversational agent (ECA), given all the

necessary characteristics explained above. An ECA is a synthetic agent that has a body and a

face, and is able to make a conversation.

Finally, it is necessary to create useful scenarios where the agents can act, showing their

personalities. It is obviously impossible to create, in the context of this thesis, a virtual world where

every single possible action in the real world is represented, therefore, the scenarios chosen, need

to be thought in detail, to allow the agents to show the qualities we are interested in, providing at

the same time, enough room for improvisation. In the scenarios, the agents will have roles, thus,

making the scenarios, role-play scenarios. Role-play is defined as the act of adopting a role [34].

This role is usually composed by personality, emotions and motivations that are different from the

ones of the individual who is role-playing [34]. Improvisational dramas and free-form theatres are

examples of role-play scenarios. In this way, role-play scenarios are a good mechanism to

visualize different personalities as they provide an easy way to define characters and watch the

result of an agent playing different characters.

1.3 Thesis Structure

Given the main problem of this thesis, we will continue by explaining its details, as well as the

solution we propose and its evaluation. The remain of this thesis is structured as follows:

In Chapter 2 we review several works in the area of psychology and computer science regarding

personality. We provide a background section that explains in general, the concepts of personality

5

and emotions, necessary for the understanding of the remain of this thesis. We then proceed to

review several works in the area of computer science where personality was embedded into virtual

agents. Finally we provide a critical analysis of the implications of these works on the current

thesis.

Chapter 3 describes the conceptual model of our solution. It explains its basic concepts,

structuring and defining the architecture of an agent with personality.

In Chapter 4 we explain how the conceptual model was implemented, describing in detail the

different modules that define a personality for the agent that can satisfy the requirements specified

above.

Chapter 5 is the chapter of the evaluation of our work. Here, we explain the methodology used to

test and validate our architecture, as well as the results obtained. Finally, we draw conclusions

based on the results, providing critics and comments on them.

Finally, in Chapter 6, we provide an overview of the whole thesis, revisiting the problem and the

hypothesis, and finally drawing conclusions about the work presented.

6

Chapter 2

Related Work

In computer science, many scientists have addressed the problem of creating individual agents,

thus, agents with different personalities. Although some interesting amount of work has been done,

one can say that this area is still in its initial stages.

However, to understand such works, we first need to review personality studies made in the area

of psychology, presented in the next subsection. As such, we will review the influence of Gordon

Allport [11], [12], [13] in personality theories. Then we will revise factor analyses made by

Raymond Catell [14], [15], [16] and also Hans Eysenck [17], [18]. Finally, we will address the Big

Five personality theory, describing it in sufficient detail taking into account its importance for our

work.

We will then proceed to review the research done by several groups working on intelligent agents

regarding personality. To begin, we will review the work of Hayes-Roth and Rousseau [29], [31],

[32], [34], [35], with personality traits in virtual actors. Next, we present an analyses of the work of

Rizzo et al. [37] with agents with personality implemented with the help of goals and plans. After

this, the work of Malatesta [33] in implementing personality in agents in order to express

appropriate mood is discussed. Afterwards, the emotional personification of virtual humans

expressed through animations proposed by Kshirsagar et al. [48] is analysed. Finally, the use of

emotions and personality to simulate a driver's behaviour and provide information about the best

driving assistance for different personalities is reviewed.

The last subsection provides a discussion on the previous ones, reflecting on the implications the

research presented had in our current work.

7

2.1 Background

There are obviously numerous studies in the area of personality, and different approaches to what

exactly personality is. Because personality and emotions are strongly related [26], we need both a

model to represent personality and a model to represent emotions. Regarding personality, the

approach used in this thesis makes use of a theory that uses traits, individual features represented

in an individual's behaviour, that are defined as basic units of personality - personality traits theory.

As for emotions, this thesis uses a theory that represents emotions as triggered by events, and

describe those emotions through the use of a few basic units, making it simple to represent an

individual's emotional state – OCC theory of emotions. Both of these theories are ideal to the

accomplishment of this thesis purpose because they give us patterns of behaviour we can use in a

systematic way in intelligent agents.

The next sections provide a review of some studies made in these areas.

2.1.1 Gordon Allport as the father of personality theory

Gordon Allport considers traits as being predispositions to act in a certain way, in response of

stimuli, and they are physical and biological features but can be seen through behaviour [11], as

for example, the way someone responds to a compliment. The resulting behaviour is consistent

because the traits lead an individual to perceive various situations as the same, and then react

accordingly to that perception [11]. They have different levels of importance and generality from

person to person, but they can be categorized: a nomothetic trait, which is a trait that can be

applied to everyone, and an idiographic trait, which is unique to a specific individual [12]. Also, a

trait can be cardinal – if it plays such an important role in a person‟s life as to lead him to behave

mainly in respect to that trait-, central – represents a disposition to behave but in a more limited

range of situations- and secondary – if it is less important in an individual‟s behaviour [13].

In his work, Allport paid more attention to the uniqueness of the individual than to the differences

between traits in different people. Although one‟s behaviour is supposed to be consistent, a slight

inconsistency of behaviour is to be expected because behaviour is a very complex concept and a

person can have different traits that can get in conflict with each other and are influenced by the

situation [11].

2.1.2 Raymond Cattell and factor analyses

Raymond B. Cattell worked with factor analysis - a statistical method - to determine groups of traits

that appeared to be correlated [14]. He saw these traits as the personality basic units and

8

discovered that there are 15 factors (traits) that are common to most of the personalities. As such,

through questionnaires intended to measure individual differences in personality, he found that

there were 16 factors which lead to the Sixteen Personality Factor Questionnaire [15].

Cattel also conducted several studies in the determinants and development of personality traits,

found that most of the traits are found in all ages and are stable especially as the person becomes

older [16]. However, he also recognized that a person‟s behaviour depends on a lot more than just

traits, like motivational factors and situations [16].

2.1.3 Hans Eysenck

Hans Eysenck‟s worked in the same line as Cattel, trying to identify the basic traits, their

determinants and stability also using factor analysis. He used the term type to describe the

dimension of a trait with a high and a low level, with every person having some value between the

two [17]. Three types are defined: Introversion-Extroversion, Neuroticism and Psychoticism, with

the first two appearing almost in every trait study [17]. The first one is related to differences in

sociability and impulsiveness: an extrovert is sociable, lively, active, while the introvert is reserved,

introspective, reflective; The type neuroticism relates to emotional instability, anxiety, body aches

and worrying, whilst the Psychoticism one is related to a tendency to be aggressive, cold,

egocentric, impersonal, unsocialized, unconventional and more creative [17]. Studies in this area

suggest that extroverts are less easily aroused than the introverts, needing a higher level of

stimulation to reach the level of arousal of the latter [18].

2.1.4 The Big Five

Although there were many studies regarding factors, they never seem to agree in the number of

basic traits until the five-factor model started to emerge [19]. This model proposed five traits as the

basic units of personality: Neuroticism, Extroversion, Openness to Experience, Agreeableness,

and Conscientiousness [19].

Among all the studies made in this area, the work of Lawrence Pervin is particularly relevant, as it

explains in a clear way the concept of personality. Traits are defined as being descriptors we use

to characterize someone‟s personality, even if we don‟t know how we got that perception [10], [22].

The use of traits to describe personality has been used since it started to be seen as an

independent area of psychology. Although the concept of traits as basic units of personality has

been credited and discredited several times along the way, the importance of traits has never been

disregarded [10]. It is said that psychologists agree that traits correspond to consistent behaviours,

9

representing differences among people, and therefore, they are useful as basic units of a

personality and it is necessary to find a way to discover and measure the basic traits [10], [22].

The five traits are defined as:

Neuroticism looks into adjustment and emotional instability, identifying individuals that

are prone to psychological distress, unrealistic ideas, excessive cravings or urges and

maladaptive coping responses. Individuals who score high in this trait have a tendency to

be worried, nervous, insecure, emotional and moody. The ones who score low are usually

calm, relaxed, unemotional, courageous, secure and self-satisfied [10], [19], [20], [21],

[22].

Extroversion is about quantity and intensity of interpersonal interaction, activity level,

need for stimulation, and capacity for joy. Individuals that score high on this trait have

tendency to be sociable, active, talkative, person-oriented, optimistic, fun-loving,

affectionate; the ones who score low tend to be reserved, sober, not exuberant, aloof,

task-oriented, retiring and quiet [10], [19], [20], [21], [22].

Openness to experience assesses proactive seeking and appreciation of experience for

the individual's own sake, toleration for and exploration of the unfamiliar. Individuals that

score high tend to be curious, creative, original, imaginative, untraditional, and have broad

interests; the ones who score low are more conventional, down-to-earth, not artistic, and

may have narrow interests [10], [19], [20], [21], [22].

Agreeableness looks into the quality of one's interpersonal orientation along a continuum

from compassion to antagonism in thoughts, feelings and actions. Individuals who score

high on this trait have the tendency to be soft-hearted, good-natured, trusting, helpful,

forgiving, gullible and straightforward; the ones who score low are usually cynical,

unfriendly, suspicious, uncooperative, fault-finding, ruthless and irritable [10], [19], [20],

[21], [22].

Conscientiousness assesses the degree of organization, persistence and motivation in

goal-directed behaviour of the individual. People who score high tend to be dependable,

fastidious, organized, reliable, hard-working, self-disciplined, punctual, scrupulous, neat,

ambitious, persevering; people who score low are usually undetermined, sloppy, aimless,

unreliable, lazy, careless, lax, irresponsible and weak-willed [10], [19], [20], [21], [22].

10

It is important to notice that traits correspond to dimensions, and the different poles of each trait

are characterized by opposites of the same features, like explained in Figure 2.1.

There are a few studies that support the validity of the five-factor model and among them there is

the cross-cultural agreement on factors, i.e., several languages found a way to describe the

differences among humans that represent the big five, suggesting that these five traits are present

in humans independently of the language [23]. Another fact that supports this theory is the fact that

there is an agreement between the individual's self-ratings and the ratings made by others [10].

This suggests that self-ratings are related to actual behaviour Also, there were found correlations

between traits and motives, emotions and interpersonal behaviour, more specifically, it was found

that individuals with a high value on certain traits tend to have certain motives, and the same with

emotions [10]. In terms of personality disorders, the trait factor model is also considered to have a

connection. In the field of genetics there has also been found a correlation to traits. Studies show

that the traits of an individual have a strong inherited component, providing genetic evidence of the

existence of traits [10].

An important issue concerning personality is its consistency, concerned to whether people express

the same trait characteristics over a range of situations, and stability, concerned to whether people

have the same trait characteristics over time [10]. It was found that the behaviour prediction based

on personality trait factors early theories was weak, because several other factors come to play

when it comes to reaction to situations [22], a phenomenon known as the person-situation

controversy [10]. However, studies also show that personality traits change very little over time,

due in part to the weight that genes have over them, and to the fact that people choose their lives

in accordance to their personality (for instance, an extrovert seeks people and activity, whereas an

introvert prefers to be alone in a quiet environment) [22]. The same trait is expressed differently in

various situations, and people behave more consistently in situations that are similar than in

Figure 2.1: Trait dimensions and their opposite features

11

situations that are different [10].

It is reliable to use behaviour in past situations to predict future behaviour in similar situations,

however, one needs to be careful when doing such thing, given that a person doesn't always

behave like it would be expected; trait theories have a large bandwidth – the amount of behaviour

that can be predicted-, but poor fidelity – the accuracy of the prediction-, that is, traits relate to

behaviour over a broad range of situations, but the accuracy of these predictions is not as good,

possibly because of the complexity of behaviour, the existence of factors that we cannot take into

account [10].

There is however, some disagreement regarding some aspects of this theory, for example,

regarding the constituents of a trait and its identity, i.e., what aspects of personality are included in

the concept of trait [19], [22]. Finally, there is no agreement on the number of basic traits nor on

their identity [10], [11], [14].

2.1.5 Emotions and Personality

Personality plays a very important role on the emotions we feel [26], thus, in order to create

believable behaviour, we will make use of this relationship. To do so, we need an emotion model

that can easily represent emotions in a systematic way. Nowadays, there is a wide variety of

psychological models of emotion [91], with focus on several different aspects of emotion theories.

Some, like dimension models [94], state that emotions can be classified using dimensions, for

example, the dimension arousal can differentiate emotions based on their intensities, from very

low to very high, in reaction to events; others, like the one proposed by Elkam [93], try to relate the

generation of a few core emotions with behaviour and physical features. It is accepted by this kind

of models that the wide variety of emotional states is a result of the combination of these few core

emotions [92]; some other, like the OCC theory of emotions [58], utilize human language to

address emotions, by analysing the meanings and connotations of the words humans use to

describe emotions.

Among the wide variety of emotion theories, the OCC seems like the most suited one because it is

a cognitive and lexical theory that defines and classifies a finite number of basic types of emotions.

Being a cognitive theory, it provides us with a mechanism to manage emotions as triggered by

objects or events [27], whilst being a lexical theory, it describes types of emotions with sets of

words, just like the Big Five personality theory describes traits, making it easier to relate emotions

and traits.

This theory proposes 22 basic emotion types. It emphasizes that each one of these emotion types

cannot be defined through a single word, but rather, by a set of words, given that this theory is not

12

about words and their single correspondent emotion, but instead, about the actual types of

emotion that one feels, and that can be described through a set of words. For example, the type

Distress includes all of the emotions that one feels in an unpleasant situation, like sadness, love

sickness and nervousness [58].

The 22 basic emotion types can be grouped in 5 different classes of emotions: attraction, fortune-

of-others, prospect-based, well-being, and attribution.

Next, we will present a short explanation about what each one of these classes means, as well as

which emotion types they are composed of.

Attraction Emotions. This class consists of the emotions one feels regarding the aspects of

someone or something. If the individual likes those aspects he feels the emotion Love, if

he dislikes them, he feels the emotion Hate.

Fortune-Of-Others Emotions. The fortune-of-others emotions are emotions regarding

consequences of events. Depending on the object involved and the pleasantness of the

event for the objects involved (including the individual himself), the following emotions can

be generated: Happy-for, if the individual is pleased with the event, and thinks that the

consequences of the event for the other are pleasant; Resentment, if the individual is

unpleased with the event, and thinks that the consequences of the event for the other are

pleasant; Gloating, if the individual is pleased with the event, and thinks that the

consequences of the event for the other are unpleasant; and Pity, if the individual is

unpleased with the event, and thinks that the consequences of the event for the other are

unpleasant;

Prospect-Based Emotions. This class regards the emotions that are felt when the

prospects of a certain event are relevant and after, when the event is confirmed or

disconfirmed, considering the pleasantness of the event. The emotions that are part of this

class are: Hope, if the individual is pleased with the unconfirmed event and is focusing on

the consequences of the prospects for self; Fear, if the individual is unpleased with the

unconfirmed event and is focusing on the consequences of the prospects for self. When

the event is confirmed/disconfirmed, the next four emotions can be generated:

Satisfaction, if the individual is pleased with the event, is focusing on the consequences

for self and the event is confirmed; Disappointment, if the individual is pleased with the

event and is focusing on the consequences for self and the event is disconfirmed; Fears-

confirmed, if the individual is unpleased with the event, is focusing on the consequences

for self and the event is confirmed; and Relief, if the individual is unpleased with the event,

13

is focusing on the consequences for self and the event is disconfirmed.

Well-Being Emotions. The well-being emotions are emotions that someone feels regarding

the pleasantness of an event, when the prospects of the event are irrelevant. One can

feel: Joy, if he is pleased with the event and is focusing on the consequences for self; and

Distress, if he is unpleased with the event and is focusing on the consequences for self.

Attribution Emotions. This class of emotions, regards the emotions felt by the appraisal of

events/actions made by self or others. This group consists of the following emotions:

Pride, if the individual approves of self; Shame, if the individual disapproves of self;

Admiration, if the individual approves of other; and Reproach, if the individual disapproves

of other;

Each one of these emotions can be felt by a person at a given situation. If an event occurs, if the

individual perceives an action, if he tries to pursuit goals, emotions are generated. There are a

number of appraisal parameters that can be specified to help generate emotions, for example, the

goals have a desirability value associated with them, that define the desirability of that goal being

achieved, which can be positive (if the individual will be pleased) or negative (if the individual will

be unpleased). If the goal is actually achieved, the desirability value will be used to calculate the

emotions that can be felt and their intensities. But how to choose the emotions that are felt by

different people with different personalities? Every emotion has thresholds and decay rates. The

thresholds define the intensity an emotion has to have in order to be felt by an individual, and the

decay rate defines the rate at which the intensity of the emotion decays. Once the possible

emotions for a given situation are generated, only the ones that exceed the thresholds, are felt by

the person [58].

2.2 Personality in Intelligent Agents

Regarding agents with personality, there is a wide variety of approaches, whether in terms of

personality theories or in architectural strategies. Let us name a few examples of the variety of

systems: Loehlin implemented ALDOUS [42], a model that uses three personality traits to simulate

the reaction of the system to an object. The characteristics of the object and the number of times

the system has encountered the object also affect the reaction. Carbonell uses personality traits

derived from goals in the POLITICS system [41]. As an example, an ambitious individual considers

the acquisition of objects and status as more important than the average individual. This work

does not present, however, a well-structured model of personality. Hovy [45] uses a natural

language generation system to explore personality traits. The OZ project [40] experiments with

14

different situations, emotions and personality traits to address the issue that an individual may

behave differently under the same conditions. This project is of great importance on the

introduction of believability of agents in the community. Nass et al. [43] proved that users react to

machines simulating dominant and submissive personalities as if they did have them. Julia [53], a

character interacting through a text-only, virtual reality system, who expresses personality and

emotions. The ALIVE project at MIT [47] is a goal-driven project that focus on emotions, which

proved that people enjoy more to interact with characters that express emotions.

Below there is a review of the most relevant systems to our work, their architectures,

implementations and results.

2.2.1 Cybercafé application

Rousseau and Hayes-Roth have a vast work on personality in synthetic agents. In [29], they

combined elements of psychology, artificial intelligence and theatre to study the use of personality

by autonomous agents, able to act and improvise their behaviour while interacting with a user.

They showed how these agents can, through personality traits, create performances that are

believable, theatrically interesting and entertaining to the users without being predictable. These

agents are like synthetic actors who improvise their behaviour in order to meet the goals of the

characters they portray.

They start with the assumption that personality is composed by a set of psychological traits. These

traits, according to them, are patterns of behaviour which are recognizable through an individual's

actions, emotions and interpersonal relationships, and do not change much over time. Also,

another perspective on personality is described, as social learning theories of personality are

briefly discussed. Regarding such theories, they affirm that a personality changes over time by

experiencing situations that are learning experiences, and that the actual behaviour of an

individual is given by his history and the situation itself [30].

After explaining the personality assumptions they will be working on, they describe some notions

on theatre, as this is the context of their system. According to them, in theatre, personalities are

used to create coherent and interesting characters. Theatre also makes use of plots - sequence of

actions involving a group of individuals - , roles – classes of characters whose typical behaviour is

known to the audience -, directors – shape the story and the characters into a production plan –

and actors – an actor has to show, using voice and body, the personality of the character he or she

represents in a consistent and believable way, and hide his or her own personality in order not to

influence the play. In [29], a different form of theatre was used, called improvisation, in which the

15

actors generate a story through spontaneous interactions with other actors and just following

instructions from a high level scenario or the audience. In the area of improvisation, [35] makes

reference to the work of Walker et al. [36] who developed a linguistic style improvisation model for

computer characters able to generate speech acts, and with is it was possible to create characters

that can improvise a speech that follows personality linguistic styles.

To accomplish improvisation, Hayes-Roth et al. [34], [35] created a set of synthetic actors, able to

generate an improvised story through actions, by interacting with each other and the users, and

following directions they receive.

The authors distinguish between two types of synthetic actors: autonomous agents and avatars.

The autonomous agents act based on the scenario constraints and the other actors, but deciding

on their own what action they will perform (directed improvisation). An avatar also receives

directions from the scenario and the other actors, but presents the possible actions for the user to

chose and them performs the selected action. These two types of actors can improvise their

behaviour according to their personality, even the avatar as the user chooses the action but it is

the avatar that chooses the way to perform that action.

Moreover, autonomous agents are agents that play characters defined in terms of personality

traits. The actions they play are defined in relation also to personality traits and moods. Such

actions, in addition to the characters, if sufficiently interesting, create an interesting story. As such

the authors tried to choose characters and plots that provoke a reaction in the audience and they

claim that this can be accomplished by using just a small number of personality traits. Another

feature that helps creating interesting scenarios is unpredictability. The unpredictability of the

actor's performance can be accomplished by introducing unexpected events in the plot or by

letting the actor perform some actions that do not reflect its personality. To do this, probability is

introduced and the actions an actor can perform have a probability factor, the actions reflecting the

personality traits having a higher probability for that actor than the ones that do not.

The characters‟ structure

The synthetic agent's mind is responsible for making the decisions, has knowledge that is updated

based on inferences and perceptual inputs. It is also the mind that makes the body act changing

the world; a body, which is responsible for representing the actions of the actor through text or

animation, and perceives the world; a mind-body interface, that links the mind and the body and

mediates their interactions. The avatar also has a user interface, responsible for showing the

options of actions to the user and giving the answer back to the mind-body interface so that the

body can act.

The knowledge in the agent's mind at the beginning of the interaction includes: the personality,

moods and interpersonal relationships of the character; the actions the actor can perform in a

16

hierarchy, characterized by personality traits, and reactions to actions performed by other actors;

the possible states of a performance defining the scenario; the world the agent performs in,

including objects, characters and places; the scripts used to follow the scenario.

– Personality is represented through the use of trait theories and social learning theories, as

previously mentioned:

To express the trait theories, the authors used a system to quantify numerically a

trait, from -10 to 10, for example a character with a value of 10 in activity means

that he is very energetic. Also the authors define a structure which correlates the

traits to other components of the model, like the actual status, for example. If a

character has a correlation factor of 0.7 between confidence and status, then it

means that his actual value of confidence is 0.7 * value of status in the current

context. This way, the personality traits are no fixed.

To express the social learning theories, they use a structure that can define a new

value for a trait if certain circumstances that change the behaviour occur. For

example, for the trait friendliness, if the character has level of anger higher than 5

(a condition), then new value of friendliness is 8. These structures also have a

priority attribute which defines which structure defines a more important situation,

in case more than one structure is defined for the same personality trait.

State machines are the part of the mind of the synthetic actor that are responsible for

telling the actual state of a part of the scenario and the next possible states. For example

the state where the character is standing can be transited by the action "walking" which

will lead to the state where the character is walking, or by the action "sit down" which will

lead to a different state

The script is composed by commands and abstract directions. The former are given by the

actor to the animator in order for it to display the actions being performed. The latter are

high-level actions that an actor must perform throughout the play.

The actions in this approach are chosen or disregarded by an actor in a certain situation

because they have a set of relevant conditions that need to be fulfilled by events in order

for the actions to be used. Then, each action has several sub-actions, called low-level

actions, which constitute several ways to perform that action. Each of these low-level

actions has annotations regarding personality traits. The actor looks at these annotations

and compares them to the personality traits of the character he is portraying, choosing the

one that has more of these traits in common. Weights are introduced to specify the

importance of each trait to a specific action. For example, if we want to say that the action

of smiling reflects more the friendliness than the confidence we can out a weight of 0.8 in

the trait friendliness and a value of 0.2 in the trait confidence in the action specification.

17

Moods are defined as emotions triggered by events and separated into two categories: the

self-oriented moods, which are not directed toward other agents, and the agent-oriented

moods, which are, even though they do not make a statement about the agent's

relationship.

Attitudes are defined as being the essence of an interpersonal relationship, that can vary

or not over time, having as examples, the status, the degree of sympathy and the trust.

All of these features are put together generating the character‟s behaviour: the mind of an agent

adds the perceptions given by the body to the situation model, identifies behaviours that match the

current context of the scenario and chooses the possible actions of those behaviours regarding his

personality traits. An autonomous actor will then calculate the probabilities for each action, choose

the most probable, and then send the commands of that action to the body. An avatar doesn't do

the last steps: he shows the possible actions he found to the user, waits for a selection from him

and then sends the commands of the action to the body. After this, the body has to express the

actions of the actor through an animator (either graphic or text animator), choosing randomly

through several ways of representing the same thing, in order to make the play more interesting.

At the beginning of a session, the actor determines the social situation (if he or she is in an group,

alone or with another character), calculates the value of each attitude and agent-oriented mood,

based on the current social situation, and determines the value of each personality trait using the

default value, a correlation to other components or an effective situation. Also, every time the value

of one component changes, the value of the related components has to recalculated. The

character has reactions to other characters actions, and the moods and attitudes can change in

these situations, and therefore the actor has to recalculate the related components. Another

situation where the actor has to recalculate the values of the components is when the social

situation changes.

Tests and results

In order to test this approach, the authors used an application of the Computer Virtual Theatre

platform [31]. They implemented a version of the Cybercafé, which provides two autonomous

synthetic actors with the role of waiter and costumer, and an avatar with the role of costumer

playing in a scenario. Through the avatar, the user can interact with the other characters, by

pressing buttons corresponding to the possible actions of his avatar.

These characters were defined with different personalities. The personality traits the authors

defined were self-confidence, activity and friendliness, and the moods were happiness (self-

oriented mood) and anger (agent-oriented mood). They also used attraction and status as

attitudes. With these components they defined ten different personalities with some of the traits

18

being independent and some being related to other aspects (like mood for example).

The first set of tests only tested the behaviour regarding personality traits. Moods and attitudes

were not included. The results were as expected: the actor behaved in accordance to his

personality most of the time but sometimes he acted in a slightly different way.

The second set of tests was made with moods and attitudes included. The results were also the

expected, for instance, the character with a selective personality, always showed anger when he

did not like the nasty agent which is consistent because the former character performed upsetting

actions more often.

The authors also performed user tests, with three different sets of personality for the characters.

After the interaction, users answered a few questions. With the answers, the authors were able to

identify which areas are more successful and which areas need improvement:

The personality recognition is easily done when moods and attitudes are not involved, and

more difficultly done when they are, probably because the interaction was too short.

In terms of believability of characters, the answers revealed that people usually find

characters believable when they have a consistent personality and are not too extreme.

Users have more problems when a character's behaviour changed over time, but again

this might be because of the short length of the interaction.

In the personification by the user test area, the authors find out that most of the people

consider the other characters' personality when they are choosing the actions for their

avatars, but disregard the avatar's personality, reacting instead with their own personality.

People react to characters as they would react to humans, having sympathy for friendly

and polite characters and disliking those who are nasty or rude.

The authors also found out that this application has potential for entertainment, as different

emotions like joy, surprise and curiosity were felt by the users.

With their work, the authors showed that that synthetic actors behave with respect to their

personalities by performing actions showing relevant characteristics most of the time. The value of

the personality traits can vary according to moods and attitudes.

It showed that this social-psychological model is robust, expressive and entertaining, providing a

good tool to create characters with flexible personalities which allow believable behaviours in a

given context.

2.2.2 Intelligent agent's processes

Rousseau did a similar work independently, with a focus on the processes of an agent, such as

19

perceiving, reasoning, learning, acting, etc. [32]. For each process, he defined a level of inclination

(tendency) and an aspect of focus and rated these along dimensions of personality traits. For

example, in the reasoning process, an actor can have high or low inclination corresponding to

being rational or silly; It can be reasoning in undesirable effects, facts or desirable effects

corresponding to being pessimistic, objective or optimistic. Rousseau claims that with the 16

dimensions of personality traits presented, one can build very rich and interesting personalities.

The actions performed by a character follow abstract rules that function as guidelines of action

expressing its typical behaviour, but these abstract rules are related to personality traits, so each

character can act differently in the same context, even though his personality will be respected.

The combination of several of these rules create the resulting typical behaviour of a character.

For example the typical behaviour of a waiter that has as personality traits realistic, insecure,

introverted, passive and secretive will be that he will do and say as little as he can.

Therefore, the possible interactions between the characters can be very interesting and

unpredictable even more because they will react to each other's personalities.

Regarding moods an interpersonal relationships, abstract directions are used to describe

preferable behaviours, for example if a character is tired he prefers to move slowly or to do

nothing.

2.2.3 Goals and Plans to represent Personality

Rizzo et al. [37] addressed the subject of personality in agents but used a different approach. They

focus on goals and plans underlying personalities to define believable agents, in order to create

complex personalities and behaviours

In [37] it is claimed that believability is present in a character if it can express thoughts, emotions,

desires and attitudes towards others, as being part of a personality. With these believable agents,

it is possible to interact on several levels, with several applications like human-computer

interaction, interactive entertainment, education and artistic domains. In order to do this, the

behaviour of such agents cannot be pre-defined, but autonomous and flexible. The most important

features of a believable agent are the emotions, expressed by physical and behavioural features,

and the personality, expressed by emotions and behaviours or by varying the traits or other

features of a personality [37].

Architecture

One of the approaches for creating personalities of believable agents is to associate behaviour

20

typical of each type of personality to emotions [2], [38], for instance, a brave personality will have

an angry behaviour if confronted with the emotion of fear. Other approaches include the

manipulation of personality traits as in [29] reviewed above, and the use of pattern attribute-value

pairs in which the attributes (for instance, intelligence) are quantified by numerical values [28]. The

approach presented in [37], uses goals to define personalities. It works on the basis that

personality is a pattern of coherent behaviours and interactions with the environment, a result of

the individual's pursuit of a set of goals. Each chosen personality needs a set of goals that

characterize it and some preferences in order to choose the plan to achieve the goals.

The goals are represented in a goal tree, after Carbonell [41], and each one of them has a general

type of action associated. Also, to each goal it is assigned a priority over the others and different

personalities will have different priorities for the goals. They chose the following goals from the

psychologically-based taxonomy developed by Ford and Nichols [44]: resource provision, like

providing approval and support and not providing selfish nor uncaring behaviour; Material gain, like

increasing the amount of money; Social responsibility, like complying to social roles and

conforming to social rules; belongingness, like caring about friendships, intimacy, and feeling in a

social community; image, like pursuing social approval and compliments about intelligence, skills

and honesty; entertainment, like searching for exciting situations and avoid getting bored.

In their architecture [37], to achieve the goals, an agent needs to have preferences related to

personality in order to choose between the possible ways of acting; two agents with the same

prioritized goal but different personalities can act differently to achieve the goal; the actions an

agent can perform are related to the goals they help to achieve but they also differ on the way they

affect other goals, therefore there can be several actions to achieve the same goal, although each

one of them has a different effect on the world, possibly getting in conflict or helping to achieve

some other goal; the prioritization of goals is combined here with preferences to decide which

actions should be performed.

This approach disregards some personality aspects like the criteria for goal activation and

achievement, and the goal-orientations which define general styles, related to personality, of

pursuing goals.

Tests and Results

In order to test the approach, the authors chose a specific behaviour, the help-giving behaviour, for

their agents, varying their personalities but always concerning the same behaviour. The reasons

for this choice include the fact that this is a very complex human behaviour, which implies some

interesting reasoning skills about the others and their motives, with bases on social relationships

like cooperation that can be useful for multi-agent systems. Also the authors have worked before

21

with help-giving strategies and motivations.

They chose altruist, normative, selfish and spiteful as the personalities that influence the help-

giving used to test the approach. The altruist agent wants to help others even if it is not good for

him or her; the normative one will help others if it does not imply breaking norms or if it complies to

some norms; the selfish one will only help others if it provides him some advantage; the spiteful

one wants to make fun of others, so it can interfere with the others' plans, refuse to help or play

tricks on them.

Each personality has some goals with the respective priorities associated. The authors claim that

in order to make a character more believable, the personalities have to be strong, and this is the

base on which they chose the goals to characterize the personalities.

An altruist agent has high priority for resource provision - because he wants to give others what

they need -, belongingness – because this goal helps to characterize the benevolence and

kindness of a character -, and image – because he wants to be appreciated for his good actions.

He has a low priority for the goal of material gain.

An agent that is normative will have a high priority on social responsibility – he cares about norms

and rules -, and image – he likes to be appreciated for his accordance to the norms.

A selfish agent has high priority for the goals of material gain – he wants possession-, and image –

because he wants others to like him so he can have more power, and the others' positive

evaluations make him feel good. He has low priority for the goals of resource provision – he

doesn't care about others unless he can get something from it-, and social responsibility - he is

willing to break rules in order to get what he wants.

A spiteful agent has a high priority on the goal of entertainment – he likes to play and play tricks on

others -, and a low level on social responsibility – social rules do not matter -, resource provision –

he likes the others to be in trouble so he will not help them -, and image – he does not care about

the image that the others have on him nor their evaluations.

As an example of verbal interaction between these agents and a human agent, it is used a

situation where the human agent asks a believable agent help to clean the dishes. The altruist will

do it, the normative will not do it because it is the human agent's turn, the selfish will do it for

money and the spiteful will say that he will do it but he will not in order to get the human into

trouble.

The tests made on this approach are text interactions between a human user and virtual agents

and are done with Prodigy, the planner, and RAP, the executer of the plans. When planning,

Prodigy makes use of functions and control rules, to select the choices in the current state, that

are in accordance with the personality of the agent, defined by the goals and preferences. These

plans are then given to RAP which executes them based on their priorities. It selects a goal,

22

defined as a task, through selection constraints and heuristics, and executes the methods

associated with that task, that are applicable in the given context and in accordance to the

personality- which consists of a plan. Therefore, in this system, different behaviour emerge for

different personalities in two levels: the planning by Prodigy creates plans related to the priorities

in a personality and then RAP executes the methods that are more appropriate for that personality.

A future development of this work is the modification of the agents' personalities according to the

feedback coming from interactions between the agents and the users.

2.2.4 Agent Personality Traits in Virtual Environments Based

on Appraisal Theory Predictions

Another approach to create Embodied Conversational Agents with personality is presented in [33].

It presents a model that generates emotions based on personality and mood. For the model of

emotions it used a simplified version of the mentioned OCC model of emotions. The simplification

consists of making the emotion generation solely based on the process of appraisal of events, and

on considering only the expressive part of the emotional reactions, i. e., the emotional state is

solely expressed, and processes like coping are not considered. This expression of the emotional

state is realized through actions/expressions and these are mapped with the use of personality. To

do this, they attribute certain personality traits to specific behaviours.

The expression of behaviour is made through facial and body expressions. In regards of facial

expressions, the predictions of K. Scherer‟s component process model theory - predicts facial

deformations in regards of the appraisal process of the stimuli presented to an individual, that

elicits emotion [91]- are mapped into MPEG-4 facial animation parameters. Regarding body

expressivity, the 6 dimensions of expressivity discussed above are used: Overall activation, Spatial

extent, Temporal Fluidity, Power/Energy and Repetitivity.

Architecture

The user interacts with the agent in a scenario, choosing actions to perform, and the agent reacts

to those actions. The input is the appraisal that the agent makes of the user. The architecture of

the character's mind is modelled through Finite State Machines, where the nodes are the states of

mood, the input is the appraisal that the agent makes of the user's action, and the output is the

expression of the agent. While in a positive mood, if the appraisal of the user's action is positive,

the output is a positive expression, and the agent remains in a good mood; if the agent makes a

negative appraisal, the output will be a negative expression and the resulting state is a bad mood.

23

The appraisal of the user's actions is made using the agent's personality, assigning different

probabilities, depending on the personality, to mood changes and action appraisal, for example, an

extrovert agent is more likely to appraise positively a neutral approach made by the user, than a

neurotic agent.

The test scenario is intended to be simple at first, and the results will be used to improve the

mapping of the emotions to expressions in order to create a more natural and believable agent.

2.2.5 A multilayer personality model

Kshirsagar et al. [48] describe a model for the personification of virtual agents. Personification

refers to the attribution of human-like characteristics and the representation of those personal

characteristics in a human-like form [48].

The Personification Model

This work's purpose is to define a personification model at the emotions level, i.e., to design minds

of virtual humans that can evolve their emotions over time, during a dialogue, proposing a layered

approach consisting of Personality, Moods and Emotions. The big five factor model is used for

personality, the OCC model is used for emotions, expressions are modelled using the 6 basic

facial expressions proposed by Ekman [49]. The mood layer is influenced by the personality and

the emotions layers. Personality influences mood changes which in turn influence the emotions

that are being "felt" by the agent. The five dimensions are mapped into good and bad mood, for

example, an agent with a high level of neuroticism is more prone to bad moods than to good ones,

and in contrary, an agent with a low level of neuroticism is more prone to good moods than to bad

ones. Then, the moods are modelled into emotions. Positive emotions constitute good moods

while negative emotions constitute bad moods. Finally, the emotions are mapped into the 6 basic

facial expressions. There is also mention to a neutral mood, which will lead the virtual human to

express emotions with a lower intensity and to not change expressions easily.

Process Overview

To accomplish this architecture, they use a Bayesian Belief Network, stating that it is the most

appropriate model when dealing with uncertainty [48].

The system makes use of MPEG-4 Facial Animation Parameters and the real-time facial animation

system using FAPs. The chat-robot ALICE [50] is used to process the input text which contains

emotional tags with probabilities. This is defined through Artificial Intelligence Mark-up Language

24

(AIML), which is a XML based language used for dialogue. These probabilities are then used by

the Bayesian Belief Network of personality which in turn updates the mood. Depending on the

specified personality, the probabilities of that personality influencing emotions and mood, and the

text response, the resulting mood is the one with the highest probability. After this, the emotional

state is updated. This update is dependent on the previous emotional state, the response

generated by the AIML and the mood updated by the personality layer. Probabilities are also used

here, specifically by defining some transition rules, for example, an agent with a bad mood has

more probability of changing his emotions to a negative one than to a positive one. After

calculating the probability of the next emotional state, the emotions that are higher than a

predefined threshold are selected and expressed. The synchronization module is responsible to

update the expression of the emotional state in a realistic way, which includes generating the lips

movement according to what the Text-To-Speech engine is producing. The result is a virtual

emotional human with a personality model that can that can express realistic behaviour during a

conversation.

2.2.6 Emotion and Personality in Driver Assistance Systems

Reichardt [54] worked with virtual agents that assist humans while driving. The success of this

assistance depends on the ability of the system to adjust to the driver's emotions and behaviours

while monitoring the driver's behaviour, calculating the adequate moment to warn or inform.

Therefore, in [54] the author constructed a system that simulates a driver's behaviour in order to

be used by assistance systems as training data. Ideally, a driver would be observed for some

moments in order to parameterize the assistant system accordingly to his personality.

To model the emotions, it is used the OCC theory of emotions. Personality is said to influence the

emotion generation of a person while she is driving, therefore, they put personality into play with

the use of the Big Five personality theory.

Three aspects of the model are defined: the adaptation of the OCC model to the driving task, for

example, the emotion of fear will be more intense if getting out of the road due to a curve has a

strong possibility and a low desirability, or if the curve is more close to the car; the integration of

personality in the model of emotions, like making emotions more intense for agents with a high

value of the Openness to Experience trait; and the incorporation of mood, for example, for an

agent in a bad mood the elicitation of the emotion Joy will be lower than for an agent in a good

mood [54].

Initial test showed that the emotion reactions of the system were acceptable [54].

25

2.3 Work Implications

The previous sections summarized several works in the areas of psychology and computer

science related to personality in synthetic agents. Let us see how we can take advantage of the

literature above to construct our work.

The Big-Five personality trait theory, is a simple and well structured way of defining, and

representing personalities, as it provides a structured definition of personality. The five traits

addressed in this theory can be easily understood and an individual's behaviour can then

correspond to combinations of those same five traits. So, although traits cannot explain fully an

individual's personality and correspondent behaviour, because there are other factors playing

along, the trait theory provides a simplified and complete enough method to understand and,

therefore, create personality.

The research by Rousseau and Hayes-Roth [29], [31], [32], [34], [35], are particularly important in

the sense that they provide a nice example of how traits can be implemented in agents. By

defining a character's personality and the actions he can perform with values for each trait

dimension, they provided a simple way to represent different personalities through different

behaviour. Moreover, the use of weights associated with traits is a powerful mechanism when

dealing with the variability of influences present between different traits and behaviour Also, their

work with the virtual theatre gives a nice mechanism to deal with role-play and improvisation since

it uses only directions of what the characters are supposed to do but leaves several options on

how they should do it using random factors and of course the personality traits. Plus, the user tests

they performed provide a useful way to evaluate a work like this.

A different approach to personality in synthetic agents is explored in the work of Rizzo et al. [37]. It

proves that goals and plans can be used to represent a character's personality in an efficient way,

by attributing specific behaviour (personality) to the pursuit of each goal, which is helpful to us

given that the agent we are creating also acts based on goals and plans.

The work of Malatesta et al. [33] is related to ours in the sense that it shows how personality can

be used to create different expressions of behaviour A personality trait is linked to mood/emotions

and behaviours and these behaviours are expressed through the use of certain expressivity

parameters, those of which can be used to differentiate personalities. In this thesis, we address

these expressivity parameters when expressing our agent's actions through the body part, relating

these parameters to personality.

26

The works of Kshirsagar et al. [48] and Reichardt [54] are important because they show how the

Big Five personality theory can be linked with the OCC model of emotions. Even though the five

dimensions of personality are mapped into specific moods, which in turn influence the agent's

emotions, it gives an insight on how both the Big Five and the OCC theories can be joined to

create behaviour dependent on personality.

27

Chapter 3

An Architecture for Agents with Personality

This chapter addresses the general concepts that guided the design of synthetic agents with

personality. It also details the conceptual model for such agents. A synthetic agent with personality

is seen as a deliberative emotional agent that has a reasoning system and is capable of speaking

and expressing emotions through body and facial expressions. The architecture of our agent is

schematized in Figure 3.1.

Figure 3.1: Architecture of a Synthetic Agent with Personality

28

Our architecture consists on a mind, responsible for processing the input from the world, appraise

that input and reason about it, and a body, responsible for receiving the output of the reasoning

system and express it visually. The reasoning leads to an emotional state which in turn, leads to

coping strategies and planning, to deal with the event perceived (the input). This planning is made

through goals and intentions, i. e., the agent selects goals to pursuit based on his or her emotions

and the state of the world, and after creating the respective intentions to achieve those goals,

constructs a plan to achieve the most important intention. The result is a plan made with a

sequence of actions, which are then sent to the body part of the agent. This part is responsible for

the visual expression of the actions the agent is performing, and includes a set of parameters used

to adjust the bodily and facial expressions performed. These actions can in turn change the world

that the agent is living in, creating more goals for the agent to pursuit. Either way, the agent will try

to achieve his goals (either new or old ones) as long as there are enough conditions for those

achievements.

We describe the agent's architecture in detail starting with the agent's mind, how personality can

be embedded in the agent's mind, describing the three levels used to accomplish this: the

emotions level, the planning level and the coping level. Next, it is explained the interface system, i.

e., how the agent expresses his actions to the user, defining the module that is used to represent

personality at the expression level.

3.1 Personality and emotions

Personality is said to influence in a major way, the way one feels [26]. Therefore, it makes sense,

when creating an agent with personality, to pay special attention to the emotions he or she “feels”

when he or she is given different personalities. Therefore, we have mapped personality traits into

emotions, so to create a relationship between the two. To do this we used the 22 emotions of the

OCC theory of emotions and the Big Five personality theory, both explained in the Background

section of Chapter 2, and related them by establishing influences that traits have over emotions.

This influence is actually translated into a definition, i. e., the personality traits define the

predisposition that the agent has to feel certain emotions, as well as their intensity, given that

personality pays such a special role on the emotions one feels. The mapping can either increase

or decrease the tendency for a given emotion and its intensity.

When looking at this mapping, it is important to remember that each trait is a continuous between

two poles of the same dimension, and therefore, an adjective used to characterize a pole can have

its antonym characterizing the other pole, i. e., if the trait Agreeableness is characterized by the

noun friendliness, it is easy to conclude that someone with a high score of Agreeableness is

friendly and an individual with a low score of Agreeableness is not friendly, thus, unfriendly.

29

Next, the relationships between the five personality traits and the emotions are exposed for every

emotion, explaining the reasons for the establishment of those specific relationships.

Joy – This emotion is defined by the trait Neuroticism given that this trait characterizes an

individual with negative emotions, that worries a lot [66], depressed [19]. This traits

contrasts emotional stability with negative emotionality such as feeling anxious, sad and

tense [22], which according to the OCC theory of emotions, are emotions that fall into the

scope of Distress [58], the opposite of Joy. Therefore, an agent with a high score on

Neuroticism, will have the tendency to feel less Joy and an agent with a low score on

Neuroticism will have the tendency to feel more Joy.

Distress – similar to the previous emotion, this one is defined by the trait Neuroticism, also

because it is a trait that reflects on less positive emotions and worrying [66], depression

[19], and contrasts emotional stability with negative emotionality such as the emotions of

anxiety, sadness, tension [22], which once again, fall into the scope of the emotion

Distress, according to the OCC's cognitive structure of emotions [58]. Therefore, an agent

with a high score on Neuroticism, will have the tendency to feel more Distress and an

agent with a low score on Neuroticism will have the tendency to feel less Distress..

Happy-for – Here, Agreeableness is the dictator of how a individual can feel this emotion.

According to [22], this trait characterizes someone who is altruist and affectionate. [19]

describes individuals with a high score on this trait as being sympathetic. All of this

adjectives are necessary for one to feel happy-for another, therefore, the more the agent is

agreeable, the more he can feel this emotion, and the less he is agreeable, the less he

feels Happy-for another.

Resentment – similar to the previous one, this emotion is defined by the trait

Agreeableness - altruism, affection [22] and sympathy [19] are all characteristics of an

agreeable person, therefore, an agent with a high score on Agreeableness, will have more

tendency to feel less Resentment, and an agent with a low score on the same trait will

have the tendency to have a strong emotion of Resentment.

Gloating – This emotion is felt when one feels good about something that is bad for

another [58]. Therefore it is defined by the level of Agreeableness given that individuals

with a high score on it are altruist, affectionate, not narcissist, humble and good natured

[22], sympathetic, does not like to show off and is modest [19]. Given this information,

30

agreeable agents are defined as having a really weak tendency to gloat, whilst

disagreeable agents are defined as having a strong tendency to do it.

Pity – This emotion is felt when one feels good about something that is bad for another

[58]. Agreeableness describes people with affection, soft-hearted [22] and sympathetic

[19], thus, it makes sense to map high scores on the trait Agreeableness to high intensities

of Pity, and low scores to low intensities.

Pride – In the literature it is found that there are two traits that influence the emotion of

Pride: Conscientiousness and Extroversion. Both of them are described with the noun

assertiveness, although there is a stronger link with Extroversion [10], [19]. Research

shows that there is a strong relationship between assertiveness and pride [67]. So, agents

who score high on these traits will have more tendency to feel pride than agents who

score low. Worth of notice is the fact that Extroversion plays a more important part on the

definition of this emotion than Conscientiousness, therefore, the agents will have to be

more influenced by the former than by the latter, regarding this emotion.

Shame – This emotion is defined by two traits: Openness and Extroversion. Having wide

interests and being unconventional [19], untraditional, liberal [66], it is easily concluded

that individuals with a high score on Openness to Experience attribute less

blameworthiness to actions and events and more praiseworthiness, given that they

engage in a lot of experiences, good or bad and like to experience the most they can get

in life [10], characteristics that however, have a weak link to shame. On the other hand,

the continuum Introversion-Extroversion regards shyness-careless [19], timid-bold and

inhibited-spontaneous [66], which are all synonyms, and according to [68], shyness and

shame have a direct correlation. Once more, the Extroversion will play a more important

role in the definition of the intensity of this emotion for the agent given that, as was said

before, it is more related to shame than openness to experience. Concluding, agents with

a high level of Extroversion and Openness to Experience, will have less tendency to

experience the emotion Shame than agents who have a high score, who will have more

tendency to feel the same emotion.

Admiration – Openness to Experience and Agreeableness have a saying when it comes to

the generation of this emotion. The former is however, weakly related as it is only linked

with admiration by presenting individuals that prefer variety, and are liberal [66], which

although provide a certain tendency for the feeling of admiration for the others' actions, are

31

merely superficially related to that emotion. Agreeableness on the other hand, has a

strong influence on this emotions given its communal orientation, and affection [22],

tender-minded, appreciative and praising feelings and actions [19], lenient, agreeable and

gullible individuals [66]. Thus, agreeable and liberal (with a high level of Openness to

Experience) agents will be able to feel more Admiration than disagreeable and

conventional ones.

Reproach – similar to Admiration, Reproach is defined by Openness to Experience and

Agreeableness. The former influences this emotion because it characterizes individuals

with narrow interests, traditional values, prefer routine [66], and once again, these only

provide a weak influence on the emotion. Agreeableness, which the low pole is described

by an antagonist orientation [22], individuals that are stingy, fault-finding, quarrelling, stem

[19], critical, antagonistic, disagreeable, and cynical [66], has all the adjectives necessary

to the conclusion that it is the most responsible for how the agent feels the emotion

Reproach. Concluding, agents with a low value of Agreeableness and Openness to

Experience will be more prone to feel Reproach then agents with a high value of those

same traits.

Love – It is easy to conclude that Agreeableness is the trait that will influence this emotion.

The continuum from compassion to antagonism in thoughts, feelings and actions [10], the

individuals being described as tender-minded, sympathetic and helpful [19], having

affection, and friendliness [22], it all indicates that agreeable people show more Love than

disagreeable people, and therefore, so do our agents.

Hate – This emotion is clearly based on the levels of Agreeableness: the continuum from

compassion to antagonism in thoughts, feelings and actions [10]. The tender-mindedness

and sympathy [19] of individuals with a high score on this trait, and the lack of affection,

unkindness, unfriendliness, cold attitude [22] of individuals with a low score, plus the two

poles being further characterized by ruthlessness-soft heartedness, Callousness-

sympathy, irritability-good nature [66], makes it simple to deduct that agreeable agents

have to have very low intensity Hate emotions, whilst disagreeable agents have to have

strong Hate emotions.

Satisfaction – Defined solely by Neuroticism with poles such as secure-insecure, self-

satisfied-self pitying [66]. A neurotic person is self-punishing [22], and given that she gives

more importance to negative emotions [10], in case of a confirmed, pleasant event, the

32

person won‟t feel as much satisfaction as a non-neurotic. Thus, our neurotic agents will

feel less Satisfaction than the non-neurotic ones.

Fears-confirmed – Defined solely by Neuroticism with poles such as at ease-nervous,

relaxed-high-strung [66]. A neurotic person is self-punishing [22], and given that she gives

more importance to negative emotions [10], in case of a confirmed, unpleasant event, the

person will feel much more the emotions Fears-confirmed than a non-neurotic one. Thus,

our neurotic agents will have more proneness for feeling the emotion Fears-confirmed

than the non-neurotic ones.

Relief – Neuroticism is the trait that defined this emotion as it characterizes people that get

upset easily, cannot accept bad things without bragging or complaining and are self-

punishing [22]. Given that they give more importance to negative than to positive emotions

[10], in case of a disconfirmed fear for an event, they won‟t feel as much relief as a people

who are not neurotic. Thus, our neurotic agents will be less prone to feel Relief than the

non-neurotic ones.

Disappointment –Neuroticism is the trait responsible for this emotion. People that get

upset easily, cannot accept bad without bragging or complaining and are self-punishing

have a high score on Neuroticism [22]. Given that they give more importance to negative

than to positive emotions [10], in case of a disconfirmed pleasant event, the person will

feel much more disappointed than others. Thus, our neurotic agents will have more

proneness to feel the emotion Disappointment than the non-neurotic ones.

Gratification – Conscientiousness affects this emotions, as it is said in the literature that

this trait delays gratification [22], provides achievement striving [19] which indicates that

non-conscientious people feel gratification more and so they don‟t put as much effort on

the tasks and goals given that they feel gratification with just a small amount of

accomplishments. On the other hand, the low pole of Neuroticism leads to self-satisfaction

and calm [66], and the high pole to depression, not contentment, self-punishment [19] and

negative emotionality [22], indicating that neuroticism influences gratification in the bad

way. With this information it is concluded that the agents we are creating have to have

weak Gratification emotions if they are neurotic, and strong Gratification emotions if they

are not.

33

Remorse – Openness to Experience is part of the Remorse emotion generation given its

continuum describes with features such as incuriosity-curiosity, adventuress-dare, liberal-

conservative [66]. Individuals who score high on it are also described as open-minded

[22], giving more importance to the experiment of a lot of things [10], giving less

importance to conventional values [19], thus less repressive and feeling less remorse

about their actions and the others‟. Neuroticism plays a slightly more important role on the

creation of this emotion because of its features like depression (not contentment) and self-

punishment [19], leading its owners to be more self-satisfying-self-pitying, Calm-worried,

self-conscious [66], and with an intense negative emotionality [22]. For these reasons, the

agents we are creating are very prone to feel Remorse if they are less open and more

neurotic, and are barely prone to the experience of this same emotion if they are not

neurotic and very open.

Gratitude – This emotion is strongly affected by Agreeableness as it needs some of the

qualities of affection [22], appreciation, warmth, praise, friendliness [19], humbleness and

gullibility [66] to be felt. Therefore, our agents feel Gratitude with more probability and

intensity if they have a high rate on Agreeableness than if they have a low rate on the

same trait.

Anger – Agreeableness is the most influencing trait on the way one feels this emotions

given its continuum defined by affectionate individuals [22] that can be defined such as

irritable-good natured, ruthless-soft hearted, rude-courteous [66], unfriendly-friendly,

unkind-kind and cruel [19]. On the other hand, Neuroticism concerns negative

emotionality, tension, not taking it easy not relaxing [22], people who are temperamental,

high-strung, with angry hostility, irritable [19], so it also has to have an influence on the

way we feel the emotion Anger. However, since this emotion in this case is directed

towards others, the Agreeableness trait in our model has more influence on the emotion

than Neuroticism. In the end, agreeable and non-neurotic agents will have a very high

threshold for Anger, meaning that they will less likely feel it, than non-agreeable, neurotic

agents, who will have a strong tendency to get angry.

Hope – Hope is defined by the level of Neuroticism an individual has. This is due to the

dimension Non-neurotic-Neurotic being described with individuals that are secure-

insecure, hardy-vulnerable [66], fearful [19] and with the tendency to have negative

thoughts about things [10]. Our agents with personality, feel more Hope if they are not

neurotic than if they are indeed neurotic.

34

Fear – Fear is defined by the level of Neuroticism an individual has, just like the emotion

Hope, being the opposite of it. Again, this is due to the dimension Non-neurotic-Neurotic

being described with individuals that are secure-insecure, hardy-vulnerable [66], fearful

[19] and with the tendency to have negative thoughts about things [10]. Our agents with

personality, feel more Fear if they are neurotic than if they are not.

Moreover, Extroversion is related to positive emotionality, as people with a high score on

Extroversion pay more attention to positive emotions than to negative emotions [22]. Also, some

studies have said that extroversion is related to the level of arousal and the need for stimuli,

stating that extroverted people need more stimuli to reach a certain arousal level than introverts

[10], meaning that introverts feel emotions more intensely than extrovert with the same amount of

stimuli. For these reasons, agents with a high score on Extroversion have this trait influencing

negatively all emotions, i. e., extroverted agents will have higher thresholds for every emotion.

To summarize what was just described, Table 1 shows the 22 emotions and their influencing traits.

However, these relationships need to be put in a more concrete and systematic way, that can be

used by our system. To accomplish this, for each trait, a specific mapping to every emotion is

made, defining this way, the exact influence that a specific value of a trait has on the value of an

emotion. But first, it is necessary to establish how emotions are represented and to do this we can

look again at the OCC theory of emotions described in Chapter 2, and represent emotions as

having a threshold and a decay rate. With these two variables, we can easily make a link between

traits and emotions, simply by defining these two parameters for every emotion, in terms of the

agent's personality traits, and accordingly to the relationships described earlier in this section. For

every emotion, the threshold and the decay rate are defined by the traits that affect them, through

the use of the following formulas:

𝑡𝑕𝑟𝑒𝑠𝑕𝑜𝑙𝑑𝐸 = 𝑖𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒(𝑇,𝐸)5𝑇=1 (1)

𝑑𝑒𝑐𝑎𝑦𝐸 = 𝑖𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒(𝑇,𝐸)5𝑇=1 (2)

where the variable T represents the five traits numbered as: Openness to experience (O) = 1,

Conscientiousness (C) = 2, Extroversion (E) = 3, Agreeableness (A)= 4, Neuroticism (N) = 5.

35

Openness to

Experience

Conscientiousness Extroversion Agreeableness Neuroticism

Joy Low Strong

Distress Low Strong

Happy-for Low Strong

Resentment Low Strong

Gloating Low Strong

Pity Low Strong

Pride Low Strong

Shame Low Strong

Admiration Low Low Strong

Reproach Low Low Strong

Love Low Strong

Hate Low Strong

Satisfaction Low Strong

Fears-confirmed Low Strong

Relief Low Strong

Disappointment Low Strong

Gratification Strong Low Strong

Remorse Medium Low Strong

Gratitude Low Strong

Anger Low Strong Medium

Hope Low Strong

Fear Low Strong

Table 3.1 – Influence of the Big Five personality traits on the 22 emotions of the OCC theory.

In the beginning of the agent's life, his personality traits scores are used in the formula above for

every emotion E, defining the corresponding threshold and decay rate.

36

Function influence(T, E) is defined as follows:

𝑖𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒 𝑂, 𝑒 =

0.3 × 𝑓1 𝑠𝑐𝑜𝑟𝑒 𝑂 , 𝑒 = 𝑆𝑕𝑎𝑚𝑒,𝑅𝑒𝑝𝑟𝑜𝑎𝑐𝑕

0.3 × 𝑓2 𝑠𝑐𝑜𝑟𝑒 𝑂 , 𝑒 = 𝐴𝑑𝑚𝑖𝑟𝑎𝑡𝑖𝑜𝑛

0.5 × 𝑓1 𝑠𝑐𝑜𝑟𝑒 𝑂 , 𝑒 = 𝑅𝑒𝑔𝑟𝑒𝑡

0, 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒

(3)

𝑖𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒 𝐶, 𝑒 =

0.3 × 𝑓2 𝑠𝑐𝑜𝑟𝑒 𝐶 , 𝑒 = 𝑃𝑟𝑖𝑑𝑒

0.5 × 𝑓1 𝑠𝑐𝑜𝑟𝑒 𝐶 , 𝑒 = 𝐺𝑟𝑎𝑡𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛

0, 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒

(4)

𝑖𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒 𝐸, 𝑒 =

0.7 × 𝑓1 𝑠𝑐𝑜𝑟𝑒 𝐸 , 𝑒 = 𝑆𝑕𝑎𝑚𝑒

0.7 × 𝑓2 𝑠𝑐𝑜𝑟𝑒 𝐸 , 𝑒 = 𝑃𝑟𝑖𝑑𝑒

𝑓3 𝑠𝑐𝑜𝑟𝑒 𝐸 , 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒

(5)

𝑖𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒 𝐴, 𝑒 =

𝑓2 𝑠𝑐𝑜𝑟𝑒 𝐴 , 𝑒 = 𝐿𝑜𝑣𝑒,𝐻𝑎𝑝𝑝𝑦 − 𝑓𝑜𝑟,𝑃𝑖𝑡𝑡𝑦,𝐺𝑟𝑎𝑡𝑖𝑡𝑢𝑑𝑒

𝑓1 𝑠𝑐𝑜𝑟𝑒 𝐴 , 𝑒 = 𝐻𝑎𝑡𝑒,𝑅𝑒𝑠𝑒𝑛𝑡𝑚𝑒𝑛𝑡,𝐺𝑙𝑜𝑎𝑡𝑖𝑛𝑔

0.7 × 𝑓2 𝑠𝑐𝑜𝑟𝑒 𝐴 , 𝑒 = 𝐴𝑑𝑚𝑖𝑟𝑎𝑡𝑖𝑜𝑛

0.7 × 𝑓1 𝑠𝑐𝑜𝑟𝑒 𝐴 , 𝑒 = 𝑅𝑒𝑝𝑟𝑜𝑎𝑐𝑕

0.9 × 𝑓1 𝑠𝑐𝑜𝑟𝑒 𝐴 , 𝑒 = 𝐴𝑛𝑔𝑒𝑟

0, 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒

(6)

𝑖𝑛𝑓𝑙𝑢𝑒𝑛𝑐𝑒 𝑁, 𝑒 =

𝑓1 𝑠𝑐𝑜𝑟𝑒 𝑁 , 𝑒 = 𝐽𝑜𝑦, 𝑆𝑎𝑡𝑖𝑠𝑓𝑎𝑐𝑡𝑖𝑜𝑛,𝑅𝑒𝑙𝑖𝑒𝑓

𝑓2 𝑠𝑐𝑜𝑟𝑒 𝑁 , 𝑒 = 𝐷𝑖𝑠𝑡𝑟𝑒𝑠𝑠 ,𝐷𝑖𝑠𝑎𝑝𝑝𝑜𝑖𝑛𝑡𝑚𝑒𝑛𝑡,𝐹𝑒𝑎𝑟𝑠 − 𝑐𝑜𝑛𝑓𝑖𝑟𝑚𝑒𝑑

0.7 × 𝑓1 𝑠𝑐𝑜𝑟𝑒 𝑁 , 𝑒 = 𝐻𝑜𝑝𝑒

0.7 × 𝑓2 𝑠𝑐𝑜𝑟𝑒 𝑁 , 𝑒 = 𝐹𝑒𝑎𝑟

0.5 × 𝑓1 𝑠𝑐𝑜𝑟𝑒 𝑁 , 𝑒 = 𝐺𝑟𝑎𝑡𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛

0.5 × 𝑓2 𝑠𝑐𝑜𝑟𝑒 𝑁 , 𝑒 = 𝑅𝑒𝑔𝑟𝑒𝑡

0.1 × 𝑓2 𝑠𝑐𝑜𝑟𝑒 𝑁 , 𝑒 = 𝐴𝑛𝑔𝑒𝑟

0, 𝑜𝑡𝑕𝑒𝑟𝑤𝑖𝑠𝑒

Function influence(t,e) corresponds to the function that defines which traits influence which

emotions, being merely a translation of what is repre sented in Table 1. As we can see in the

formulas above, there are different weights for the contribution of each trait to the threshold/decay

(7)

37

rate of different emotions, representing the fact that some traits have more influence than others

on the way we feel emotions, just like we saw earlier in this section. The chosen weight values

were chosen based on what can be found in the literature about the differences between the

different influence of different traits. Finally, the actual value of the threshold/decay rate, defined

with values between 0 and 10, according to a certain trait (before the weights), is defined through

functions f1(x), f2(x) and f3(x):

The formulas created to implement the relationship between emotions and traits, were defined in a

way that they allow for a synthetic agent to perform a behaviour that complies the one described in

the literature.

Given all this information, the agent's we are creating are able to feel emotion depending on the

personality we attribute them.

3.2 Personality and Planning

Another way we can incorporate personality in synthetic agents is through manipulating the

process of planning process. As said before, our agents have goals, intentions and plans. Upon

the selection of a possible achievable goal, a plan is created, composed by actions, in order to

achieve the goal. The point of relevance for personality here, is that the agent has many goals he

can pursuit at a given moment in time, and therefore can construct several plans to achieve every

single one of them. Imagine, for example, that the agent has the goal to go to the cinema. He can

simply go, making this a plan with one action, or he can first get money, see if it is better to take

his car or ask for a ride, etc. As a result, plans can have several measurable characteristics: the

number of actions, the importance of the plan, the number of things that executing that plan can

deal with, the number of things that need to be accomplished in order for the plan to succeed, and

its success probability.

Among the five traits, Conscientiousness is the one that was found to be more relevant to this

process. Conscientiousness describes several characteristics of the human being that are

naturally part of the planning process, such as the degree of organization, persistence and

motivation in goal-directed behaviour. People who score high tend to be fastidious, organized,

38

reliable, hard-working, scrupulous, ambitious, persevering and deliberative; people who score low

are usually undetermined, sloppy, aimless, lazy, careless, negligent and thoughtless [10], [19],

[20], [22], [66]. When it comes to plan, our agents also have these characteristics affecting the

process.

Next, we describe our model for how Conscientiousness can influence the planning process:

Number of actions: the number of actions constituting a plan is important for the agent.

Characteristics of someone with a low level of Conscientiousness are for example, lazy

and thoughtless as opposed to someone with a high level of Conscientiousness, who is

fastidious and hard-working [19], [66]. Thus, our choice for this influence resides in that an

agent with a low level of Conscientiousness will prefer plans with a small number of

actions to a plan with a lot of actions, and of course, an agent with a high level of

Conscientiousness will not give such a big importance to how many actions a plan has,

being the result of a plan the important thing for him.

Number of open preconditions: the number of open preconditions of a plan refers to the

number of facts that need to be true, before the plan can start to be executed. This line of

thought falls in the same one as the previous point, therefore, the influence will be defined

as the same, that is, an agent with a low level of Conscientiousness will prefer plans with a

small number of open preconditions to a plan with a lot of open preconditions, and an

agent with a high level on this trait will not give much importance to this number when

choosing the best plan for him.

Probability of success: a plan has the purpose to achieve something, to satisfy some goal,

however, this is not always accomplished, as sometimes, things on which the plan

depends on are changed, or simply when the results of some actions involve a certain

degree of uncertainty. In this context, plans have success probabilities associated with

them. This is the probability of a plan actually accomplish what it was made for. For

someone who is very conscientious, setting up goals and persist until they are achieved is

a very important quality [10]. These are people who have the features of being goal-

driven, fastidious, reliable, ambitious, persevering, thoughtful and deliberative [10], [19],

[66], making them people who give a big importance to achieving their goals, and finishing

what they proposed [22]. As a result, our agents interpret a plan's probability of success in

different ways. An agent with a high score on the trait Conscientiousness puts a bigger

weight on the probability of success when choosing a plan for execution than an agent

39

with a low score, given that, as was said before, for the first agent it is more important to

succeed when executing a task, than for the second.

An agent prefers a plan to another based on the value of an heuristic: the value is calculated for

every plan, and the agent chooses as best plan, the one with the lowest value. The mechanism to

choose the best plan makes use of the variables of planning described above and complies the

following formula:

For each plan p, an heuristic h is calculated. Variable stepsSize corresponds to the number of

steps of the plan (number of actions in order to achieve the goal, for instance), variable

openPreconditions corresponds to the number of pre conditions that need to be fulfilled before the

plan can be execute, variable PT regards the number of threats that exist by going forward with

the plan p and probability corresponds to the probability of success of that same plan.

As we have seen, personality affects differently the importance each person gives to these

variables, and that difference is implemented in our agents with the use of weight functions.

Function wSS(x) gives a weight to the number of steps in the plan, wOP(x) balances the

importance of the number of open preconditions and wP(x) weights the value of the probability of

the given plan, where score(C) is the score that the agent has on the Conscientiousness trait.

These weight functions are defined as follow:

Once more, these function were constructed in a way that complies to what it is said in the

literature about the influence of the trait Conscientiousness in these planning attributes.

40

The heuristic described above is essential for the personalization of the planning process: all of

the plans the agent constructs are put through this heuristic which attributes different importance

to different features of the plan, according to personality, and the plan with the lowest value is then

selected for execution. This way the agent can select the plan that most suites his personality, his

way of acting.

3.3 Personality and coping

Personality has been shown to influence the coping process of an individual [61], [62], [63], [64],

[69], and the results of this process [70], [71]. Research within the big five personality traits theory

has shown that these traits can provide information regarding the individual's coping process [61],

[72], [63], [69]. However, there is not an agreement on which traits influence coping nor how they

influence it. The most studied traits are neuroticism and extroversion, however there has been

found a relation between coping and Openness, Agreeableness, and Conscientiousness [61], [62],

[63], [69]. Moreover, several researchers have noted that failing to control for the other personality

dimensions when examining a specific personality dimension may lead to spurious results [73],

[78], [63], [69]. There is also the opinion that it is only Neuroticism that has a direct influence on

the coping process and the other traits come to play solely by their correlation with Neuroticism

[87], [82], [62], [63]. Also, there is evidence that coping is a result of an interaction between the

individual and his or her environment [79], which leads to the hypothesis that the discrepant results

found in all these studies are due to a relationship between personality and the environment that

was not taken into account, i. e., the same trait may have its influence depending on the context

[63], [86], [76], [83], [80].

Below, there is a brief description of what the literature says about the influence of each one of the

big five traits in the coping process.

Neuroticism is the trait that has the most consensus about its influence on the choice of coping

strategies. This trait manifests its influence by leading an individual with a high value of this trait to

use more passive or emotion-focused coping strategies like escape avoidance, self-blame, wishful

thinking, and relaxation [62], [74], as well as interpersonally antagonistic means of coping such as

hostile reactions, catharsis, confrontative coping [61], [71], [63], or interpersonal withdrawal [63],

than problem focused strategies [61], [75], [71], [72], [63], [64]. Studies showed that an individual

with a high value of this trait, will choose coping strategies associated with negative outcomes [77],

[81], [85], and will have more negative outcomes even when he applies supposed better coping

strategies [70]. In [80], a high value of Neuroticism is associated with the coping strategies of

41

interpersonal withdrawal, escape avoidance, self-blame, and support seeking.

Regarding Extroversion, studies have shown that it influences the coping process by leading

extroverts to engage more in problem-focused coping [78],[62], [64], and less in supposed

maladaptive forms of emotion focused coping such as self-blame, wishful thinking, and avoidance

[78]. These individuals tend to use more support seeking [73], [61], [78], [69], positive thinking or

reinterpretation [62], [69], and substitution and restraint [62]. On the other hand, some studies

have raised some doubts on such theories, by experimenting with this trait while holding the

effects of the other traits constant. Basically, these studies did not find a relationship between

these trait and problem-focused coping [78], [63] or adaptive forms of emotion-focused coping

[61], [63]. Those higher on Extroversion were significantly more likely to report engaging in

compromise and self-blame [80].

As for those with a high value of Openness, some research shows that they are more likely to

engage in positive reappraisal [63], [69], and to plan their coping [69]. However, some studies

have found no significant relations between this trait and coping [78], and others have found it to

be only a weak predictor of coping [62], [84]. Individuals higher on Openness were significantly

less likely to report using distancing than were those lower on it [80].

Agreeableness has been associated with coping strategies that protect social relationships like

seeking support [78], [63], [84] and avoiding confrontation [63]. Those high on this trait are more

likely to use positive reappraisal and problem solving [84], [69] and tend to avoid emotion focused

coping strategies such as self blame, avoidance, wishful thinking [78], or disengagement [69] than

those low on this trait. Those higher on Agreeableness were significantly less likely to report

engaging in self-blame, with a non-significant trend toward being less likely to report engaging in

escape avoidance [80]. However, all the evidence that relates this trait to coping has not been

enough to support these relations [78], [84]

There is some evidence that Conscientiousness influences coping [84]. People with a high value

of this trait tend to use more active, problem-focused strategies [78], such as planning, problem

solving, and positive reappraisal [69], and less likely to employ emotion-focused coping strategies

such as self blame [78], [63], or disengagement [69]. Finally, those higher on Conscientiousness

were significantly more likely to report engaging in relationship-focused coping, compromise,

problem solving, and support seeking [80].

Given all this information, it is complicated to decide how exactly personality traits influence the

42

process of coping, and even more, if they actually influence it. Because Neuroticism is the only

trait that has enough support by researchers on the matter of influencing coping, we chose to

influence coping strategies with Neuroticism wherever it was possible, following a positive

interpretation of the literature regarding this trait. This means that if there is some research saying

that this trait influences a coping strategy, we assume it does. Given all of the controversy

regarding the other traits, we chose not to consider them in the coping process.

Looking at our agent's possible and implementable coping strategies, we see that there is one

which is said to be influenced by neuroticism: denial, also known as wishful thinking.

In our model, this coping strategy is applied when, for example, there are certain goals that are

being threatened by an active plan. In this case, the agent denies those threats, going forward with

the plan. To make this dependent on the personality, this strategy is only chosen by the agent,

depending on his level of Neuroticism: if it is high, the agent applies denial/wishful-thinking, if it is

low, the agent does not.

3.4 Personality and expressivity

In order to express the agent's behaviour, it is necessary to create a body that can perform the

expressive actions. We are interested in a body that can also express personality, and to do this, it

has to be able to express various features that compose one's personality, like for instance,

emotions. Therefore, our agent needs to be able to express emotions visually, as well as body

movements that go along with the action the agent performs at a given moment, which, if analysed

efficiently, can be manipulated to express different personalities. This can be accomplished with

the use of facial expressions and gestures. Research shows that one's facial expressions

influence the judgement that other people make about one's personality [88]. This means that

facial expressions bias the perception of personality, even if this perception is inaccurate. Also,

several studies have shown the importance of facial expressions in the perception of the emotional

state of the person who is showing them [95], [96], [97], [98]. In fact, the distinction between

emotional state and personality is not clear when it comes to the physical expression of emotions,

with the result of leading other people to perceive one's personality according to one's emotional

displays [65]. Moreover, people make accurate assumptions about the emotional state of one

another through the observation of bodily signals, even though most of the times, they do not

realize they are doing this based on those signals [65]. Because an individual's facial expressions

can help others to predict the individual's behaviour, and his behaviour is influenced by his

personality traits, there is a link between his facial expressions and the attribution that the others

make of his personality traits [88].

Regarding other bodily expressions, several studies have shown the relationship between the

arms position, facial expressions, vocal expressions and posture expressions, and personality

43

traits [99]. Despite the studies mentioned, there is still not a clear definition of the link between

concrete personality traits and bodily expressions [88], therefore, the assumptions made to

manipulate them to express different personalities have to be solely superficial.

Given the information above, it is easily concluded that in order to represent personality in a

synthetic agent that can perform facial expressions and gestures, it is essential to analyse in which

way these facial expressions and gestures can induce personality perception. One of the ways to

address this is through the expression of emotions, the agents' emotional state, which in turn

reflects the agent's personality (if, of course, the relationship between emotions and personality is

well defined). The other way is through the manipulation of some expressivity parameters in facial

expressions, gestures and head movements. By adjusting those parameters, it is possible to make

the same action look slightly different, representing, this way, different personalities performing the

same action. Personality is then introduced at this level through the expression of emotions (which

as we have seen earlier are an expression of personality) and through the diversity of ways that

the agent can express the same gesture, manipulating the agent's expressivity parameters.

We consider six expressivity parameters in order to vary the expression of the same behaviour.

Next, we will describe those parameters, and how we can use them to express personality.

Spatial extent – this parameter has to do with the amount of space required to perform an

expression, translated in for example, the extension of the arms and the amplitude of

eyebrow raise. Extroversion has to do with expressiveness, and the more extroverted a

person is, the more expressive she is [10]. Wide and ample movements reflect on a bigger

expressiveness and also, research has shown that a lot of expressiveness provokes more

empathy on the person who is watching who in turn perceives the other as extroverted

[99], [65]. Given this information, our synthetic agent with personality expresses the

actions with a lot of spatial extent if he is extroverted, and with a few if he is introverted.

Temporal extent – this parameter regards the amount of time spent to perform an

expression, translated in for instance, quick or slow movements. Impulsive movements are

linked to Extroversion [99] as is the amount of energy of a person [22] which reflects itself

in a high speed of movement. Moreover, rapid facial expressions of smiling were linked to

openness to experience [kinematical]. Given that although there is reference in the

literature to the influence of Openness to Experience on the speed of facial expressions,

the influence documented is not as big and general as the influence made by

Extroversion, we decided to assign high scores in Extroversion to a short temporal extent

of movements in our agents, and also attributed a small influence on the short temporal

44

extent to high scores in Openness to Experience,

Fluidity - this parameter concerns the smoothness of movements, translated in, for

example, graceful or jerky movements. It is easily observed that the excess of enthusiasm,

related to Extroversion, leads people to move faster (as mentioned above), and with less

coordination, performing more sudden movements. This is also true for Neuroticism, but in

this case it is related to stress and nervousness, which lead to more jerky movements. As

a result, our agents have a high value of Fluidity if they are not extroverted nor neurotic,

and a low value if they are.

Power - this parameter regards the intensity of an expression, translated in, for instance,

strong or weak movements. The power of an expression also has to do with the amount of

expressiveness, which as we have seen, is linked to Extroversion, being either positive or

negative expressions [99]. Also, studies have shown that dynamic movements are scored

by the observer as extroverted [65], and because of all of this, our agents with personality

have a high power value for their movements if they are extroverted, and a low power

value if they are not.

Overall Activity - this parameter regards the total amount of activity during a whole set of

movements. When speaking, the agent can perform head, facial and body movements at

the same time, and this parameter is used to define the overall quantity of movement

performed across that set. As we have seen before, the amount of expressiveness is

strongly related to Extroversion. Also, some studies have found a correlation between

overall expressivity and positive emotionality which is a characteristic of extroverted

people [18]. Therefore, agents with a high level of Extroversion have also a bigger amount

of overall activity than agents with a low level of Extroversion.

Repetitivety - this parameter regards the repetition of certain movements. Here, the

literature does not make any reference regarding personality, but one can assume that the

repetitivety shows a certain instability to the observer, for example reflecting nervousness,

and as those a characteristics of neurotic people [10], our agents have a high value of this

parameter if they have a high score on Neuroticism, and a low level if they don't.

Once more, it is necessary to translate these observations to actual relationships that define the

framework and can be used in a system. We defined the possible values of these parameters as

being from -1 to 1, representing the lowest and the highest level of a parameter respectively, and

45

created the following formulas to map the conclusions made above into concrete values.

Where score(x) is the agent's x trait score, and x can be E (Extroversion), O (Openness to

Experience) and N (Neuroticism), used in functions expressivity(x) and expressivityRev(x) which

were created to translates values from 1 to 5 (the traits) into values from -1 to 1 (the expressivity

parameters) and are defined as:

As we have seen before, some traits have more influence on certain parameters than others, and

that is what the w1 and w2 variable are for. w1 corresponds to 1 meaning that the trait has its

whole value influencing the parameter (sometimes in conjunction with another trait), and w2

corresponds to 0.5, meaning that the trait value is not as important to the definition of the

parameter as another, or that the value does not play such an important role on the attribution of a

value to the parameter.

3.5 Summary

In this chapter we have provided an overview of the architecture of our agents with personality. It is

composed by the agent's reasoning system, which gathers the knowledge the agent has about the

46

world, his or her goals, and reasons about the best way to achieve those goals. Based on the state

of the world, emotions which are generated accordingly to world events, are biased by the agent's

personality. The expressive system, accomplished with an Embodied Conversational Agent,

responsible for expressing facial expressions, gestures, postures and speech, that are generated

with respect to what the agent's emotional state and reasoning outcome. We have seen how

personality comes to play into all these parts by influencing the appraisal process of the agent,

manipulating the emotions that an agent feels when appraising events, by influencing the coping

process and the planning of actions and by adjusting the agent's expressivity parameters. We

explained how every emotion is influenced differently, depending on the agent's traits. Formulas to

map the relationships found in the literature between emotions and personality traits into actual

values were also created. We then proceeded to explain how the coping strategies applied are

also different according to the agent's traits, as well as the plan he chooses to execute. Here, more

formulas were introduced to accomplish the same as before: establish a way to translate the links

found in the literature into concrete values.

Finally, concerning the agent's body, we have shown how six expressivity parameters can be

adjusted by the agent's personality, in order to more realistically reflect it through facial and body

expressions, and once more, provided some formulas that help us find values that represent what

is said in the literature, to use later in the implementation of our agent.

47

Chapter 4

Implementation

This chapter describes the implementation of the architecture defined in the previous chapter,

explaining the details of the solution that was used to accomplish the proposed conceptual model.

First it is described how FAtiMA's system [39] was linked with GRETA's [51] in order to provide us

with a mechanism that enables us to develop an agent with a mind and a body, and then how the

personality module was created and connected with these two, in order to create synthetic agents

with personality. Finally, the mechanism of creating a personal agent, the authoring, is described.

4.1 FAtiMA and GRETA

Our agent's mind is created with the use of FAtiMA, and the body with the use of GRETA. FAtiMA

reasons about the world and GRETA performs the actions the mind tells it to perform. The system

works like this: FAtiMA's agent chooses goals to pursuit at a given point, and then creates plans,

sequences of actions, to accomplish those goals. He then performs each single action that

constitutes the plan. This action is sent as output to the virtual world where the agent is acting.

Alongside the action the agent has to perform is the agent's emotional state at that moment, which

is composed with the emotions that the agent is feeling and their intensities. On the other hand,

GRETA works by receiving as input, a BML message containing the speech to be spoken and the

gestures/postures/facial expressions accompanying the speech. The agent's personality process

is schematized in Figure 4.1.

48

4.1.1 FAtiMA

As said before, the agent's mind is accomplished through FAtiMA [39]. FAtiMA is a system that

uses goals and plans to create deliberative agents. The agents are in a virtual world, have beliefs

about it, perceiving events and acting upon those events, trying to pursuit their own goals. After

selecting a goal, the agent creates an intention to achieve the goal. Then, a plan of actions is

created for that intention in order to achieve the goal. However, FAtiMA's process is a little more

complicated than this. Because FAtiMA‟s agents are emotional agents, emotions play an important

role in the selection and execution of goals and plans, respectively. Goals can be selected

according to emotions, for instance, imagine the goal fightback, which achieves the action of

fighting another agent that has been mean to our agent; we might define this goal to be chosen

only in the case our agent feels the emotion anger. Another way that emotions influence the

agent's behaviour is with action tendencies, which define actions that the agent needs to perform if

he is feeling a certain type of emotion with a certain intensity, given a specific event. Moreover, the

agent can appraise events. This appraisal generates more emotions, which in turn influence the

goals that the agent tries to achieve, and the actions he uses to achieve them. By creating plans,

the agent creates sequences of actions that are chosen based on, among other things, the

Figure 4.1: The personality influencing agent's acting process

49

emotions that the agent feels at that time, in order to achieve his goals. Another process in the

agent's mind is the coping process. Upon the creation of plans, more emotions are generated, for

instance, emotions of fear and hope about the failure or success of the goal he is pursuing, or

about the threats that executing a given plan will produce. The agent needs to deal with this and to

do so, he uses coping strategies. Then, the agent selects the best plan and executes the actions.

Next it is presented an overview of FAtiMA's architecture as well as its most important processes

for our work, the appraisal process and the coping process.

Architecture

The architecture consists in a cognitive layer, where the emotional planner creates emotions that

are then selected according to appraisal variables (which include the emotional thresholds and

decay rates), and a reactive layer consisting in emotional rules.

There are two types of goals: active-pursuit goals, and interest goals. The latter are things that the

agent wants but doesn't pursue, the example given being an agent wanting his football team to win

a game. Both these goals will influence the emotions that the agent generates, either by failing or

succeeding (Disappointment/Satisfaction, etc.) or simply by trying to be met by the agent

(Hope/Fear). These emotions have a set of characteristics associated with them, like arousal and

moods. The arousal is characterized by different levels. For instance, an agent experiencing very

strong emotions will have a very high level of psychological arousal. The moods are said to

influence the agent's emotional state and also to be influenced by it. For example, an agent with a

good mood is believed to experience more positive emotions than an agent with a bad mood. So,

there is a lot more to emotions in FAtiMA than just their name and intensity, and emotions easily

become a complex concept in a agent. To deal with such complexity the architecture of an agent in

FAtiMA has two processes: appraisal and coping. The appraisal process is responsible for

generating emotions that are appropriated to the context of the interaction at that moment, creating

plans according to the goals and the environment, and then create their correspondent intentions.

The intentions are assigned initial emotions, and the intention with the strongest emotion is

selected to be pursued. The coping process is responsible for adding actions to plans in order to

satisfy the intentions, according to the agent's emotions and personality. A plan is created with

actions that will lead to the desired result and the emotions and personality of the agent will

constraint these actions, for example if a plan is in conflict with another, a fearful agent will more

likely drop the plan then a courageous agent.

Appraisal process

The appraisal represents an evaluation made by the agent about events. The character's appraisal

50

to generic events is defined through a set of rules and attributions:

The emotional reaction rules represent the character's standards and attitudes and are

very tied to personality [58]. These emotional reactions define actions that the agent

performs in reaction to events on an emotional level.

The action tendencies correspond to reactive actions, that the agent performs without

reasoning.

The threshold defines the intensity level the emotion has to reach in order to affect the

agent, and the decay rate defines the time that the agent is affected by that emotion [58].

When an event occurs, a set of emotions is generated but only the emotions that reach the level of

threshold specified are added to the character's emotional state. However, this process of

generating emotions makes use of appraisal variables defined through numeric values, like the

desirability of emotional reactions, that triggers emotions as Joy (if the desirability value is positive,

meaning desirable event) and Distress (if the desirability value is negative, meaning undesirable

event); the desirability for other, that triggers emotions as Pity, Gloating, Sorry-for and

Resentment; the praiseworthiness/blameworthiness of an action, that generates emotions such as

Pride, Shame, Admiration and Reproach; the like/dislike values for the other agents and objects in

the world, that create Love and Hate emotions, and so on [46].

Coping process

In order for the agent to deal with events, he needs a coping system. FAtiMA's coping process is

defined with a few strategies: emotion-focused strategies such as acceptance, mental

disengagement, wishful thinking/denial; and problem focused strategies such as planning and

execution. These strategies are applied when the agent perceives an event, and needs to cope

with it. The problem-focused strategies are applied when the agent needs to plan his response.

The planning consists of analysing the possible actions that the agent can perform, and use them

to create a sequence that the agent will then execute, in order to achieve the desired goal. There

can be generated more than one plan, and if this is the case, the agent chooses the most relevant

according to a special heuristic. This is where the emotion focused coping comes into action. If the

chosen plan generates a strong Fear emotion, for example, by threatening some of the agent's

interest goals, the agent can decide not to execute the plan, applying the acceptance coping

strategy (accepting that the plan will not be executed due to its risk), or he can go forward with the

plan, either by ignoring the threats applying the wishful thinking/denial strategy, or by accepting the

failure of the interest goal using the acceptance coping strategy.

51

After the plan is chosen, the agent starts to executing by performing the first action of the plan. the

coping process makes sure that the plan is being executed, by updating and executing the actions.

Every action is perceived by the other agents, and this creates more emotions, generating more

action-tendencies and plans to achieve new goals.

4.1.2 GRETA

To create our agent with personality, we use GRETA's embodied conversational agent [51], [52] for

the interface, because GRETA is an agent who is able to express emotions, and also vary that

expression according to parameters. This way, the actions and the corresponding emotional state

that come from the agent's mind are sent to GRETA which then expresses them, recreating human

behaviour.

GRETA is an application that expresses human behaviour through gestures, facial expressions

and posture [51]. It is able to speak and perform the mentioned bodily expressions at the same

time, having the main purpose of simulating in a believable way the conversational human

behaviour. To do this, the agent makes use of the face, through the manipulation of eyebrows,

eyelids, mouth and cheeks; head movements, using the head direction and rotation, such as nods

and shakes; gestures, through the manipulation of arms and hands; and body posture, through the

manipulation of the upper part of the body. Finally, GRETA uses emotions to simulate a human in a

realistic way: using the features mentioned above, she is able to express a wide range of emotions

in a believable way.

Architecture

GRETA makes use of a Baseline and a Dynamicline [52]. The former is used to define the general

ways of expression that one has. For example, an agent can be usually make few gestures, and

speak with his arms close to his body, while another can make a lot of wide gestures while he

speaks. The Dynamicline is used to represent the behaviour in a more local way, that is, how an

agent behaves in a specific situation.

GRETA uses expressivity parameters in order to implement distinctive behaviour [52]. These

parameters are part of the gestures/facial expressions and are the same as the ones we are

considering for our own agent: Overall activation (the quantity of movement), Spatial Extent

(expanded or contracted movement), Temporal Extent (fast or more sustained movements),

Fluidity (continuity of movement, smooth or sudden), Power (weak or strong gestures), and

Repetitivity (rhythmic repeats of movements) [52].

52

The Behaviour Quality Computation module receives as input the baseline and the communicative

intention and/or emotional state, and generates the Dynamicline based on that input [52].

However, it is possible to do this manually, that is, define, ourselves, what facial expressions (that

will most likely be representative of the agent's emotional state), gestures and other movements

the agent does while performing.

Let us take a look at how this agent works. The agent has a baseline which indicates the basic

expressivity of the agent, i. e., how the agent usually expresses herself. When nothing is given for

the agent to do, he reacts based on this concept, waiting for an instruction. Then, actions are sent

to the agent, which include the sentences, the emotions she should express, and can also specify

the gestures she should perform. These expressions are then parameterized according to the

baseline creating the Dynamicline, i. e., if the agent performs quick expressions according to the

baseline, then she will perform the sent expression quickly according to the Dynamicline. This way,

the agent can express the same gesture in different ways.

The process consists of the following: the agent is waiting for instructions; a message is sent to the

agent containing the expressions and the speech she has to perform; the agent calculates the

resulting expressions and speech, and performs them, while she speaks.

Connecting FAtiMA to GRETA is then just a matter of translating the output of the former to the

input of the latter. It is then necessary to create a BML message containing the action the agent

wants to perform (the speech) and gestures that go along with that action, as well as facial

expressions that reflect the emotional state of the agent.

4.2 Implementing Personality

In order to implement personality it is first necessary to find a way to represent it. To do that, we

have looked at personality models with the purpose of finding how the specialist of the area

assess and represent one's personality. There are a few different models regarding the five traits,

with two of the most famous being the NEO-PI-R/NEO-FFM model [90] and the Big-Five [89]. The

differences between these two consist basically on the theoretical explanations they use to

describe the structure of the traits and dimensions. We decided to use the Big Five Inventory,

given that it is free to use. Following the notation of the Big Five Inventory [89], we define each trait

with a number from 1 to 5. This way, to attribute a personality to an agent it is only necessary to

specify his scores on each of the five traits. These scores are defined in a XML file and read by the

system in the beginning of the agent's “life”, and are specified as follow:

53

<Personality name="melancholic">

<Traits Openness="1"

Conscientiousness="4"

Extroversion="4"

Agreeableness="3"

Neuroticism="5">

</Traits>

</Personality>

This way, any personality already analysed by the Big Five Inventory can be transferred to an

agent.

Next, we describe how personality can be implemented in the four main features we have

discussed in Chapter 3 that are used to represent personality: the appraisal process, the planning

process, the coping process and the bodily expressivity.

4.2.1 Implementing personality in the appraisal process

As it was explained in Chapter 3, emotions are felt by an agent according to his personality traits.

Every agent has emotional thresholds and decay rates for each of the 22 emotions from the OCC

theory, defining the way they feel those emotions. In FAtiMA, emotions are generated according to

several appraisal parameters, and after this, their intensities are compared to the agent's

emotional thresholds. Only the emotions that have an intensity value greater than or equal the

value of the emotional threshold for that emotion will be felt by the agent. Then, the decay rate is

used to calculate the pace at which the intensity of the corresponding emotion decays. Using this

mechanism we can implement the relationship between emotions and personality traits, by

manipulating emotional thresholds and decay rates using the formulas proposed in Chapter 3. As

an example, an agent with a high score on Agreeableness will have a low threshold and decay

rate on the emotion Love, and a high threshold and decay rate on the emotion Hate.

4.2.2 Implementing personality in the planning process

FAtiMA's agents use a certain heuristic to help them decide on the best existing plan. This

heuristic makes use of a formula similar to (11) presented in Chapter 3, only in FAtiMA's case,

personality weights for each of the components are not used. It is then easy to implement the

proposed model for planning, given that the personality scores for the agent are already in the

54

system, and so, the agent only needs to access them in order to calculate the resulting heuristic

for each plan. The plan with the lowest value is the chosen one.

4.2.3 Implementing personality in the coping process

Upon the selection of a plan, FAtiMA analyses various coping strategies in a sequence, verifying if

there are any that can be applied. Among the various coping strategies that are verified, one

regards the possibility that in order to execute a plan, an agent might have to choose to let go of

some interest goals that are being threatened by the new plan, generating the emotion fear. As we

have seen before, the strategy of denial leads the agent to ignore existing threats in order to lower

his fear emotion, and it is chosen more frequently by neurotic agents than non-neurotic ones. To

accomplish this strategy in FAtiMA, we attribute a value to a control variable, denial, based on the

following formula:

𝑑𝑒𝑛𝑖𝑎𝑙 = 𝑓2(𝑠𝑐𝑜𝑟𝑒(𝑁 − 1)) (23)

where f2(score(N)) is defined as (9) and score(N) is the score on the Neuroticism trait attributed to

the agent.

Upon the analyses of the possible matching coping strategies, the agent verifies if the number of

existing conflicts generating fear is higher than denial. If it is, the agent will apply this strategy. It is

important to notice that (23) results on a low value for the denial variable is the agent is neurotic

and a high value if the agent is not. Therefore, even if there is a high number of conflicts, if the

agent is not neurotic, he will also have a high value on denial, thus, we will unlikely engage in this

strategy. When applying this coping, the agent will choose to ignore some conflicts, lowering their

probability, thus, decreasing the intensity of the emotion fear. Because a neurotic person does not

ignore every single threat, just like a non-neurotic person does not accept every single threat, so

do our agents. The number of conflicts the agent actually ignores is also dependent on his value of

Neuroticism, and is given by this formula:

n= nIgnoredConflicts− denial (24)

where nIgnoredConflicts is the number of existing conflicts and denial is the variable defined in

(23). This way an agent only decreases the probability of n conflicts, making it possible for agents

who are very and agents who are not. Imagine two agents: one has a score of 5 and the other a

score of 2 on Neuroticism. The first will have a value of 0 on the denial variable and the second will

have a value of 6. Now imagine that there are seven conflicts. Both of the agent will apply this

coping strategy but the first, the neurotic one, will lower the effect probability of the seven conflicts

55

and the second, the non-neurotic one, will lower the effect probability of only one.

4.2.4 Implementing personality in bodily expressivity

After the agent has decided what to do based on his goals, emotions, coping, etc., he sends his

actions as input to his embodied conversational part, accomplished with GRETA [51]. This input is

then translated into speech, gestures and facial expressions as we have seen before. In order to

represent the action with respect to his personality, performing it in slightly different ways, the six

expressivity parameters are defined. We've seen how they can be manipulated depending on the

personality and we have seen how these relationships are accomplished with formulas. Before

GRETA's system receives the input, it is necessary to transform FAtiMA's output to match

GRETA's input syntax. The action is transformed into a speech accompanied by gesture(s), head

movement(s), and facial expression(s). These physical expressions are defined in the system to

correspond directly to actions, so it is only necessary to transform the action chosen by FAtiMA

into the physical expressions used by GRETA. The rough XML definition of these actions is as

follows:

<face id="emotion-1" start="5.71" end="7.29" stroke="0.56">

<reference>affect=anger</reference>

<FLD.value>-0.85</FLD.value>

<OAC.value>1.00</OAC.value>

<PWR.value>0.90</PWR.value>

<REP.value>-0.10</REP.value>

<SPC.value>0.50</SPC.value>

<TMP.value>0.55</TMP.value>

</face>

The first line defines the type (facial, head or gesture movement), name and the duration of the

movement, as well as its starting and finishing moment. The reference field regards the name of

the movement. In the example above we can see that this facial expression refers to the emotion

anger so, the agent will express that emotion through his face. The next six lines define the six

expressivity parameters: Fluidity (FLD), Overall Activity (OAC), Power (PWR), Repetitivity (REP),

Spatial Extent (SPC) and Temporal extent (TMP). When creating this message, the system uses

the agent's personality traits to find values for these parameters using the formulas proposed in

Chapter 3 (15), (16), (17), (18), (19) and (20).

The message is then constructed and sent to the part of the agent responsible for the visual

56

expression of behaviour.

4.3 A practical example

Now let us provide a practical example of a scenario with an agent embedded with personality

using the architecture proposed in this thesis, describing how the agent was created and the

process that lead her to behave in the way she did. In order to create an agent it is necessary to

specify a set of parameters that define the agent's mind, like goals, actions, emotional reaction

rules, action tendencies and relationships. Obviously, it is almost impossible to define every

possible goal, action, etc., in our human world for an agent, so, this wide range of possibilities

needs to be restricted for the purpose we are creating the agent. In the case of this example, the

context of the scenario consists of a job interview guided by an agent without a personality

architecture; the interviewee is our agent with personality. Next we will describe how we can

specify all of the agent's parameters according to the context of a job interview.

4.3.1 The agent's role

FAtiMA requires the definition of an agent through a XML file that constitutes the role that the

agent will have in the scenario. This file defines some of the appraisal parameters that we have

showed to determine the generation of emotions, like the goals‟ importance of success and failure,

the interpersonal relations, the emotional reaction rules and their desirability and praiseworthiness,

as well as the actions that are performed when certain emotions occur, called action tendencies.

Below there is an example of each case as defined for our character, Lisa, the interviewee.

<Goal name="RespondCompetitivetyConfident([target])" importanceOfSuccess="6"

importanceOfFailure="3">

</Goal>

<Goal name="RespondCompetitivetyLeader([target])" importanceOfSuccess="6"

importanceOfFailure="3">

</Goal>

<Goal name="RespondCompetitivetyKind([target])" importanceOfSuccess="6"

importanceOfFailure="3">

</Goal>

<Goal name="RespondCompetitivetyFear([target])" importanceOfSuccess="6"

importanceOfFailure="3">

57

</Goal>

The importance of success and failure are representative of what the agent thinks about the

respective importance of those goals, regardless of her personality.

<InterpersonalRelations>

<Relation target="Salsa" like="6" respect="8"></Relation>

</InterpersonalRelations>

<EmotionalReaction desirability="5" desirabilityForOther="-6" praiseworthiness="3">

<Event action="SpeechAct" subject="[SELF]"

parameters="RespondCompetitivetyConfident">

</Event>

</EmotionalReaction>

<EmotionalReaction desirability="-2" praiseworthiness="-4">

<Event action="SpeechAct" subject="[SELF]" parameters="RespondCompetitivityFear">

</Event>

</EmotionalReaction>

Once again, it is important to notice that the desirability and the praiseworthiness represent the

agent's point of view regarding those actions, that is, Lisa thinks that responding to a question

about competitivity with fear is not desirable (-2) because it will probably be a negative point on the

her evaluation by the interviewer and is blameworthy (-4) representing that she does not think that

that is something to be proud of, however, responding to the same question with confidence is

desirable (5), undesirable of the other agent (-6) because it shows that she may be competition

and of course that she is confident, and she thinks that it is praiseworthy (3).

<ActionTendency action="showArrogancy">

<ElicitingEmotion type="Pride" minIntensity="1">

</ElicitingEmotion>

</ActionTendency>

This action tendency will ignite the action showArrogancy if the agent feels pride with at least the

intensity of 1.

58

In FAtiMA, this file is also responsible for defining the emotional thresholds and decay rates for

every single one of the 22 basic types of emotions that we have seen before. However, in our

model, those parameters are defined according to the agent's personality, so their definition in this

file is irrelevant.

4.3.2 Defining goals

The set of goals available in the world where the agent acts are in the goal library, which is also a

XML file. There can be numerous goals in this file, and the agent does not have to use them all,

obviously: she will only use the ones that she can. To decide which goals she can pursuit, goals

have preconditions associated that define a set of conditions that need to be fulfilled before the

agent can achieve the goal. If a certain goal has certain preconditions that the agent cannot meet

ever, that goal is never considered by the agent. As an example let us look at a defined goal:

<ActivePursuitGoal name="RespondDoubtReasonsToHireAngry([target])">

<PreConditions>

<Property name="[target]" operator="!=" value="[SELF]"></Property>

<Property name="[SELF](role)" operator="=" value="Candidate"></Property>

<Property name="[SELF](actionMoment)" operator="=" value="10"></Property>

<Property name="[target](actionMoment)" operator="=" value="11"></Property>

<EmotionCondition emotion="Reproach" min-intensity="2">

</EmotionCondition>

<EmotionCondition emotion="Distress" active="False" minintensity="3">

</EmotionCondition>

</PreConditions>

<SucessConditions>

<RecentEvent occurred="True" subject="[SELF]" action="SpeechAct" target="[target]"

parameters="RespondDoubtReasonsToHireAngry,11">

</RecentEvent>

</SucessConditions>

<FailureConditions>

</FailureConditions>

</ActivePursuitGoal>

When a first agent tells a second agent a number of characteristics that he thinks are enough to be

59

hired by a company, and the second agent responds with a comment that reflects a certain doubt

about those being enough reasons to be hired, the first agent can have an angry reply, defined by

the goal above. As we can see, it is an active pursuit goal and there are a number of preconditions

defined in it. The first states that the agent can only pursuit this goal (reply with anger to the

doubts) if the agent who expressed those doubts is not the agent herself. This prevents the agent

from replying to herself. The second precondition says that only an agent with the role of

Candidate. The third and fourth preconditions regard action moments. These are properties

defined in FAtiMA to help in the construction of scenarios that make use of plots. If we want the

agents to have a set of different goals on different moments of the story, we define action moments

to specify which goals belong to which moment in the plot. In the example above, the third

precondition states that the agent can only reply with anger if she is in the action moment 10

(meaning that she has answered about the reasons she thinks are enough to be hired by the

company), and the fourth precondition states that the agent can only pursuit this goal if the action

moment of the agent she is replying is 11 (meaning that the other agent already made the

comment regarding his doubts about the interviewee). The fifth and sixth preconditions are

influenced by personality because they regard emotions. In this case, the agent will only be able to

follow this goal if he feels the emotion Reproach with at least intensity 2 and does not feel the

emotion Distress with more than intensity 3. The values attributed have to make sense regarding

the goal they are defined for. In this case, an agent that feels too distressed is more likely to

engage in stressed/nervous/depressed behaviour, thus the limit for Distress and of course she has

to feel Reproach. It is important to notice that other emotional conditions could be defined, with the

same results, however, one needs to be sensitive and analyse the goal and situation in order to

define only necessary and sufficient preconditions.

After this, the success conditions are defined. In this case, the goal is only viewed as achieved if

the agent performs the speech act (usually one or two sentences) correspondent to replying to the

doubts with anger. There can also be failure conditions, as explained before, but in this case, no

such conditions needed to be defined.

<InterestGoal name="AvoidHumiliation">

<ProtectionConstraints>

<Property name="[SELF](humiliated)" operator="=" value="False"></Property>

</ProtectionConstraints>

</InterestGoal>

This interest goal simply defines a property that the agent wants to keep protected, in this case,

the agent does not want to feel humiliated.

60

4.3.3 Actions and inference moments

Actions are also defined in a specific XML file, and define effects/changes in the world and are

usually performed to achieve certain goals. Actions, like goals, can have preconditions, and have

also a specified effect.

<Action name="SpeechAct([target],[type],[variable])">

<PreConditions>

<Property name="[target]" operator="!=" value="[AGENT]"></Property>

</PreConditions>

<Effects>

<Effect probability="1">

<Property name="SpeechContext()" operator="="

value="#EVENT([AGENT],SpeechAct,[target],[type],[variable])">

</Property>

</Effect>

<Effect probability="0.6">

<Property name="[target](humiliated)" operator="=" value="True">

</Property>

</Effect>

</Effects>

</Action>

This action corresponds to a speech act, with only one precondition, defining as the target of the

speech act, an agent other than the agent performing the action. As effects, there are defined two:

the first changes the speech context with probability 1, meaning that it is certain that the agent

performs the event of a speech act to the specified target with the given parameters; the second is

a possible (0.6 probability) humiliation that the agent gets by saying something in an interview.

There is another type of action, called inference moment, responsible for updating the action

moments mentioned earlier and providing certain actions (like the speech act presented above)

with parameters.

<Action name="InferenceMoment01.1([ag1])">

<PreConditions>

<Property name="[ag1](role)" operator="=" value="Juri"></Property>

<Property name="[ag1](actionMoment)" operator="=" value="0"></Property>

61

<RecentEvent occurred="True" subject="[ag1]" target="[ag2]" action="Question"

parameters="greet"></RecentEvent>

</PreConditions>

<Effects>

<Effect probability="1">

<Property name="[ag1](actionMoment)" operator="="value="1">

</Property>

</Effect>

</Effects>

</Action>

In the inference moment presented above, the action moment of the agent with the role of Jury

(the interviewer) is updated to 1 if it is of 0 at that moment and if the Jury has just asked another

agent a question with the parameter “greet” (a form of speech act related to greeting other agent).

4.3.4 Relating actions to expressions

We have seen how the actions planned by the agent need to be translated to bodily expressions in

order to be expressed by GRETA. These associations are defined in a file, and relate every action

with a set of movements, as it is exemplified below.

<Association>

<ActionClass>SpeechAct</ActionClass>

<ActionSubclass>RespondBackgroundSad</ActionSubclass>

<HeadFile></HeadFile>

<GesturesFile>RespondBackgroundSad.gesture</GesturesFile>

</Association>

This association links the speech act (tag ActionClass) of answering about an agent's own

background in a sad way (tag ActionSubclass) to a gesture file (tag GesturesFile). It could also be

linked to a head movement but it was left in blank in this case, and possibly be selected randomly

by the system. The files containing the gestures simply have a link to an existing GRETA gesture

like the following:

<reference>EMOTION=SADNESS</reference>

62

This will then be included in the input sent to GRETA as explained in section 4.2.1.

4.3.5 The result

After all of these parameters are defined (including the agent's personality), it is possible to launch

the agent(s) and watch the emergent behaviour. As a result, Lisa, our agent, will answer to the

interviewer according to what is more natural to a person with her personality, her facial

expressions will reflect her emotions which will also be the ones more expected for someone with

her personality as well as the expression of her gestures. Figures 4.2 and 4.3 show the resulting

performance of Lisa with a cheerful personality (Neuroticism of 1, agreeableness of 4, extroversion

of 5) and with a neurotic personality (Neuroticism of 5) respectively.

Figure 4.2: Agent with a cheerful personality (O-3; C-3; E-5; A-4; N-1)

Figure 4.3: Agent with a neurotic personality (O-3; C-3; E-3; A-3; N-5)

63

4.4 Summary

In this chapter we have seen how we can implement the architecture proposed in Chapter 3, using

two other systems: FAtiMA [39] for the mind and GRETA [51] for the body. We have seen how we

can implement the influence made by personality traits on emotions through the manipulation of

emotional thresholds and decay rates used in FAtiMA, the influence made on the planning process

through the use of an heuristic that uses FAtiMA variables, the influence made on the coping

process by interfering in the cycle that analyses the coping strategies to apply, and the influence

made on the expression through the definition of the six expressivity parameters in BML language.

Finally, we saw a practical example of the system, describing step by step the authoring process of

our system.

´

64

Chapter 5

Evaluation

We have shown the conceptual model and implementation of synthetic agents with personality and

now we will describe their evaluation. This evaluation intended to validate the model and

implementation created, through the analysis of the perception of the personality of an agent. To

do this, we have created some user on-line tests, where the users can watch a scenario with

agents performing a situation, and then provide information about their perception concerning the

personalities of the agents. The actual parameters we wanted to evaluate were the personality

traits themselves, and to do this it was essential to test the application with different values for the

same trait. So, we created four different tests, each one with the same scenario performed with a

different personality, making the traits vary from performance to performance. If the users could

identify correctly the traits of the agents it would mean that our approach was valid and we were

able to convey different personalities through the created architecture.

The scenario chosen for the agents consists in a job interview. The agent with personality is

interviewed by another agent, and tries to answer the asked questions in a way that can lead her

to get the job, never forgetting her personality.

5.1 Test structure

The test is composed of a personality questionnaire regarding the user, a video with a

performance of an agent with a specific personality and another questionnaire, this time regarding

the agent's personality. The personality questionnaire [22], [89] used is part of the Appendix A.

Four different conditions were created, each one with a different video varying the personality

65

traits, as it was explained above. The differences between the videos are based on the differences

between the four basic personality types [100]: sanguine, melancholic, phlegmatic and choleric.

These types are described in the literature with adjectives, making it easy to attribute them actual

scores in the five traits, even though the main purpose of this choice of personalities was not to

define correctly the four types but rather, find an easy way to come up with four different

personalities. The four basic types and their translation to the BFI score are described as follows:

a melancholic is a thoughtful ponderer with a melancholic disposition, kind and

considerate, creative, overly pre-occupied, depressed. Also often a perfectionist,

unsatisfied with himself or herself. We attributed to this type the following scores:

◦ Openness To Experience: 1

◦ Conscientiousness: 4

◦ Extroversion: 4

◦ Agreeableness: 3

◦ Neuroticism: 5

a sanguine is light-hearted, fun-loving, a people person, loves to entertain, spontaneous,

leader abilities, confident and energetic. However they can be arrogant, cocky, and

indulgent. We attributed to this personality the following scores:

◦ Openness To Experience: 3

◦ Conscientiousness: 1

◦ Extroversion: 5

◦ Agreeableness:4

◦ Neuroticism:1

a choleric has a lot of ambition, energy and passion. He can dominate people of other

temperaments, especially phlegmatic types. On the negative side, they are easily angered

or bad-tempered. We attributed to this type the following scores:

◦ Openness To Experience: 1

◦ Conscientiousness: 5

◦ Extroversion: 5

◦ Agreeableness: 1

◦ Neuroticism: 3

a phlegmatic is self-content and kind, shy, lazy and resistant to change. Very consistent,

relaxed, rational, curious, and observant, has many friends, is reliable and compassionate.

We attributed to this type the following scores:

66

◦ Openness To Experience: 3

◦ Conscientiousness: 3

◦ Extroversion: 1

◦ Agreeableness: 5

◦ Neuroticism: 2

5.2 Results

The test was made available on-line so that it could reach to a larger number of users. In total we

got 46 answered questionnaires, with the age of the users who answered it ranged from 20 to 30

years old with a similar number of male and female participants. Next we will analyse the results

for each one of the five personality traits, except Openness to Experience given that it is a trait

more related to actions (painting, reading, etc.) than actual emotions and expression, for each one

of the four personality types used. The data representing the referenced mean values is available

in Appendix B.

The Melancholic character

The results of the tests with the melancholic personality show that the users identified personality

traits with the following mean values (the values between parenthesis regard the original score):

3.19 (4) for Conscientiousness, 2.42 (4) for Extroversion, 3.24 (3) for Agreeableness and 4.04 (5)

for Neuroticism. Comparing with the original scores we can conclude that except for Extroversion,

every trait was identified as being very close to the original score, differing from less than 1 point.

Figure 5.1 summarizes the results of the scores for each trait identified by the users.

67

5.2.1 The Sanguine character

The results of the tests with the sanguine personality show that the users identified personality

traits with the following mean values (the values between parenthesis regard the original score):

3.53 (1) for Conscientiousness, 4.11 (5) for Extroversion, 4.06 (4) for Agreeableness and 2.32 (1)

for Neuroticism. Comparing these results with the original trait values, we can conclude that the

traits Agreeableness and Extroversion were the ones more correctly identified by the users, and

the trait Conscientiousness was the least. Figure 5.2 summarizes the results of the scores for each

trait identified by the users.

Figure 5.1: Mean values for the four analyzed traits in the melancholic personality

68

The Choleric character

The results of the tests with the choleric personality show that the users identified personality

traits with the following mean values (the values between parenthesis regard the original score):

3.82 (5) for Conscientiousness, 3.59 (5) for Extroversion, 2.76 (1) for Agreeableness and 2.02 (3)

for Neuroticism. Comparing these results with the original trait values, we can conclude that

although the scores identified by the users were not equal to the original score, they were all

relatively close, especially when analysing the results with the use of the standard deviation.

Figure 5.3 summarizes the results of the scores for each trait identified by the users.

Figure 5.2: Mean values for the four analyzed traits in the sanguine personality

69

The phlegmatic character

The results of the tests with the phlegmatic personality show that the users identified personality

traits with the following mean values (the values between parenthesis regard the original score):

3.54 (3) for Conscientiousness, 2.51 (1) for Extroversion, 3.55 (5) for Agreeableness and 2.59 (2)

for Neuroticism. Comparing these results with the original trait values, we can conclude that the

traits Conscientiousness and Neuroticism were the ones more correctly identified by the users,

although every trait had a mean value close to the original scores, especially when applying the

standard deviation to the calculus. Figure 5.4 summarizes the results of the scores for each trait

identified by the users.

Figure 5.3: Mean values for the four analyzed traits in the choleric personality

70

5.3 Interpretation

Although the results were not completely accurate it is easy to see that all of the 4 personalities

tend to the original ones. The four classified trait scores were always similar to the ones the agent

really had, and even though the results did not match exactly, we could see that the main

characteristics that define each one of the 4 personalities were identified by the users, for

example, even though the mean for Agreeableness in the phlegmatic agent was 3.55 and the

original value was 5, we can see that this trait has the highest score in the users' classification,

and even though the mean for Extroversion in the same personality type was 2.51 and the original

score was 1, we can see that this trait has the lowest score in the users' classification, showing

that the personality perceived had the same tendencies as the real personality of the agent,

although the values were not exactly the same. This can be due to several things. First, the scores

Figure 5.4: Mean values for the four analyzed traits in the phlegmatic personality

71

attributed by us to the agent may not match the actions we introduced in the world, for instance,

defining the action of replying angrily to an interviewer may not be characteristic of a person

feeling the intensity of Reproach that we attributed to that action, therefore, our agent could have

chosen actions that did not match entirely her personality (and her emotions in this case).

Secondly, the variation of expression of the same action is essential for the perception of the trait

Extroversion (which was one of the traits with the biggest difference between original and identified

scores), meaning that the body needs to be constructed with that in mind, putting special

relevance to the variation of gestures and facial expressions, which was not the main goal of

GRETA. Moreover, some user perceptions can be biased by their own personality. Although no

significant correlation was found between the user traits and their evaluation of the agent's

personality, it was possible to notice nonetheless that some trait combinations had correlation

values much lower than others suggesting that there might be an influence of the personality of the

users and their perception of personality. More exhaustive studies in that area can be done by an

independent work in order to accurately identify these correlations, as this issue fall out of the

scope of this thesis.

Concluding, the results were positive as they showed that the traits perceived by the users were in

the same area of the original scores (for example, a high value of Neuroticism was identified as

high, even though the scores were not the same), meaning that the personalities were correctly

perceived in general terms.

72

Chapter 6

Conclusions

Personality is becoming an extremely important issue to address in the field of intelligent agents,

given the importance of personalization and individualization of systems that arises from the

globalization of this era and the need for uniqueness of the individual that comes from it. Given

this, we have proposed a personality model for a synthetic agent aiming at agents that can

demonstrate coherent and consistent personalities.

To do this, we have looked at several personality theories in order to know exactly how personality

is defined and structured and chose one theory that most suited our main objective: the Big Five

traits theory. This theory describes personality as being a set of predispositions, that one has, to

act in a certain, consistent way, and these predispositions can be defined as traits. It defines five

traits as being the basic ones: Openness to Experience, Conscientiousness, Extroversion,

Agreeableness and Neuroticism.

We have also looked at emotion theories, in particular the OCC theory of emotions, which defines

22 emotions as the basic types of emotions.

After an extensive review of the literature regarding personality in psychology and in computer

science, we proposed an architecture for our agent with personality. In this architecture we

proposed that personality should be an influence on some aspects of the agent mind and body.

The mind is composed by goals, plans and actions, and the body, by bodily expressions. It was

then explained what aspects of the agent were influenced by personality in our system: in the mind

we considered the agent's emotions, coping process and planning process, and in the body its

expressivity.

We have then related each personality trait to the 22 emotion types by establishing an influence

73

between traits and emotions. Regarding the coping process, we have made a link between

personality traits and the coping strategies chosen. As for the planning, we have modelled

personality traits to influence the selection of the best plan to execute at a specific moment.

Finally, regarding bodily expression, we have related personality traits and the way gestures and

facial expressions are performed.

To implement the proposed architecture we have used two systems already available: FAtiMA [39]

- a reasoning system - for the mind, and GRETA [51] – an embodied conversational agent - for the

body. The resulting process was the following: the agent perceives an event, updating his or her

knowledge base and generating emotions that the agent then “feels” according to his or her

personality. Also, the agent can select goals and try to pursuit them by creating plans of actions.

Upon the creation of these plans, some conflicts can arise and the agent will analyse the situation

and apply the necessary coping strategies. After the coping process, the agent chooses the best

plan for him and executes them through actions. These actions are sent to the body to perform,

which has gestures linked to actions and facial expressions linked to emotions, both of which are

parameterized in order to act in accordance with the agent's personality.

To test our approach we have created a set of 4 conditions answered by users. In each one of

these conditions, participants saw a video with a different personality type each, and the

participants were asked to answer some questions about themselves and about the character in

the video. By analysing the results we concluded that the users could identify correctly the

personalities portrayed by the agent in the video. It is important to notice, however, that this

system is very dependent of the authoring of the scenario and agent, given that, in order for

personality to be coherent and consistent, it is essential to specify the necessary goals, actions,

gestures, etc., all part of the agent's mind and body. It was not found any significant correlation

between the personality of the users and their perception of the agent's personality. However there

was minor evidence that this could be indeed related, leaving the answer for further studies.

74

References

[1] Atkinson, R. L., Atkinson, R. C., Hilgard, E. R.: Introduction to Psychology. Harcourt Brace

Jovanovich, Inc. (1983)

[2] Bates, J.: The role of emotion in believable agents. Communications of the ACM, 37(7), 122–

125 (1997)

[3] J. Rickel, W. L. Johnson; Task-Oriented Dialogues with Animated Agents in Virtual Reality;

Workshop on Embodied Conversational Characters. In Proceedings of the Workshop on

Embodied Conversational Characters; October 1998

[4] Loyall, A. B., Bates, J.: Personality-rich believable agents that use language. In AGENTS ‟97:

Proceedings of the first international conference on Autonomous agents, pp. 106–113, New York,

NY, USA. ACM Press. (1997)

[5] Ortony, A.: On making believable emotional agents believable. 189–211, Emotions In Humans

And Artefacts, Cambridge, MA: MIT Press. (2001)

[6] Moffat R.: Personality parameters and programs, In Creating Personality for Synthetic Actors,

Springer. (1997)

[7] Ryckman, R.: Theories of Personality. Belmont, CA: Thomson/Wadsworth. (2004)

[8] C. Nass, B. Reeves; Social Response to Communication Technology; Stanford University;

http://www−leland.stanford.edu/group/commdept/

[9] M. Mateas; An Oz-Centric Review of Interactive Drama and Believable Agents; Technical report

cmucs -97-156, School of Computer Science, Carnegie Mellon University, Pittsburgh; June 1997

[10] Pervin, L.: The Science of Personality. Wiley, New York. (1996)

[11] Allport, Gordon W.: Concepts of trait and personality. Psychological Bulletin, 24, 284-293.

(1927)

[12] Allport, Floyd H., Allport, Gordon W.: Personality traits: Their classification and measurement.

Journal of Abnormal and Social Psychology, 16, 6-40. (1921)

[13] Allport, G.: Becoming: Basic Considerations for a Psychology of Personality. Yale University

75

Press (1960)

[14] Cattell, R. B.: The meaning and strategic use of factor analysis. In J. R. Nesselroade, R. B.

Cattell. Handbook of multivariate experimental psychology. (2nd ed.). New York: Plenum. (1988)

[15] Cattell, Raymond, B.: Personality: A systematic theoretical and factual study. (McGraw-Hill

publications in psychology), New York, NY: McGraw-Hill. (1950)

[16] Cattell, Raymond B.: Personality and motivation structure and measurement. World Book Co.

(1957)

[17] Eysenck, H. J.: Dimensions of personality London: Routledge, Kegan Paul, p. 279-303.

(1947)

[18] Strelau, J., Eysenck, H. J.: Personality Dimensions and Arousal. Hardcover. (1987)

[19] John, O. P., Srivastava, S.: The Big-Five Trait Taxonomy: History, Measurement, and

Theoretical Perspectives. University of California at Berkeley. (1999)

[20] Howard, P.J., Howard, J. M.: The Big Five quickstart: An introduction to the Five-Factor Model

of Personality for human resource professionals. Charlotte, NC: Centre for Applied Cognitive

Studies. (1995)

[21] McGowan, J., Gormly, J.: Validation of personality traits: A multicriteria approach. Journal of

Personality and Social Psychology, 34,191-795. (1976)

[22] John, O. P., Naumann, L. P., Soto, C. J.: Paradigm Shift to the Integrative Big-Five Trait

Taxonomy: History, Measurement, and Conceptual Issues. (2008) In O. P. John, R. W. Robins, L.

A. Pervin (Eds.), Handbook of personality: Theory and research (pp. 114-158). New York, NY:

Guilford Press.

[23] McCrane, R., Costa, P.: Personality Trait Structure as a Human Universal. National Institute

on Aging. American Psychologist, Vol. 52, No. 5, 509-516. (1997)

[24] Cattell, R. B.: The Scientific Analysis of Personality. Penguin. (1965)

[25] Allport, G. W.: Traits Revisited. American Psychologist 21: 1-10. (1966)

[26] Arnold, M., B.: Emotion and personality. New York, NY, US: Columbia University Press. (1960)

[27] Lazarus, R.: Emotion and adaptation. New York: Oxford University Press. (1991)

[28] Goldberg, A.: Improv: A system for real-time animation of behaviour-based interactive

synthetic actors. In Trappl, R., and Petta, P., eds., Creating Personalities for Synthetic Actors.

Berlin: Springer. (1997)

[29] Rousseau, D., Hayes-Roth, B.: Personality in Synthetic Agents. Report KSL 96-21. Knowledge

Systems Laboratory, Stanford University. (1996)

[30] Bandura, A.: Social Learning Theory. Englewood Cliffs, Prentice-Hall. (1977)

[31] Hayes-Roth, B., van Gent, R. Story-Making with Improvisational Puppets and Actors.

Technical Report KSL-96-05, Knowledge Systems Laboratory, Stanford University. (1996)

[32] Rousseau, D.: Personality in Computer Characters. Knowledge Systems Laboratory, Stanford

University. (1996)

76

[33] Malatesta, L., Caridakis, G., Raouzaiou, A., and Karpouzis, K.: Agent Personality Traits in

Virtual Environments Based on Appraisal Theory Predictions. Image, Video and Multimedia

Systems Lab, National Technical University of Athens, PENED programme. (2003)

[34] Rousseau, D., Hayes-Roth, B.: Improvisational synthetic Actors with Flexible Personalities.

Report KSL 97-10. Knowledge Systems Laboratory, Stanford University. (1997)

[35] Rousseau, D., Hayes-Roth, B.: A Social-Psychological Model for Synthetic Actors. Report KSL

97-07.. Knowledge Systems Laboratory, Stanford University. (1997)

[36] Walker, M. A., Cahn, J. E., Whittaker, S. J.: Linguistic Style Improvisation for Lifelike Computer

Characters, Technical Report WS-96-03, Portland, OR, 61-68. (1996)

[37] Rizzo, P., Veloso, M., Miceli, M., Cesta, A.: Personality-driven social behaviours in believable

agents. AAAI Fall Symposium on Socially Intelligent Agents. (1997)

[38] Elliot, C.: The Affective Reasoner: A process model of emotions in a multi-agent system. Ph.D.

Dissertation. Northwestern University. Technical Report no.32. (1992)

[39] Dias, J., Paiva, A.: Feeling and Reasoning: A Computational Model for Emotional Characters.

GAIPS Grupo de Agentes Inteligentes e Personagens Sintéticas, INESC-ID and IST, Tagus Park.

(2005)

[40] Bates, J., Loyall, A. B., Reilly, W. S.: An Architecture for Action, Emotion, and Social Behaviour

Technical Report CMU-CS-92-144, School of Computer Science, Carnegie Mellon University.

(1992)

[41] Carbonell, J. G.: Towards a Process Model of Human Personality Traits. Artificial Intelligence

1-2: 49-74. 1980.

[42] Loehlin, J. C.: Computer Models of Personality. New York: Random House. (1968)

[43] Nass, C., Moon, Y., Fogg, B. J., Reeves, B., Dryer, D. C.: Can Computer Personalities Be

Human Personalities? International Journal of Human-Computer Studies 43: 223-239. (1995)

[44] Ford, M. E.: Motivating Humans. Goals, Emotions, and Personal Agency Beliefs. Newbury

Park (CA): SAGE. (1992)

[45] Hovy, E. H.: Pragmatics and Natural Language Generation. Artificial Intelligence 43: 153-197.

(1990)

[46] Dias, J.: FearNot!: Creating Emotional Autonomous Synthetic Characters for Emphatic

Interactions MSc Dissertation. Instituto Superior Técnico. (2005)

[47] Maes, P.: Artificial Life meets Entertainment: Lifelike Autonomous Agents. Communications of

the ACM 38 (11): 108-114. (1995)

[48] Kshirsagar, S., Magnenat-Thalmann, N.: A Multilayer Personality Model. MIRALab, CUI,

Geneva, Switzerland. (2002)

[49] Ekman, P.: Emotion in the Human Face. Cambridge University Press, New York. (1982)

[50] ALICE A.I Foundation: http://www.alicebot.org/

[51] Pelachaud, C.: Multimodal Expressive Embodied Conversational Agents. University of Paris 8,

77

LINC IUT de Montreuil. (2005)

[52] Mancini, M., and Pelachaud, C.: Implementing distinctive behaviour for conversational agents.

University of Paris 8, INRIA - France. (2006)

[53] Mauldin, M.: Chatterbots, TinyMUDs, and the Turing Test: Entering the Loebner Prize

Competition. Proceedings of the Twelfth National Conference on artificial Intelligence (AAAI-94),

Seattle, WA. (1994)

[54] Reichardt, D. M.: Emotion and Personality in Driver Assistance Systems. Proceedings of the

4th Annual Meeting on Information Technology and Computer Science: 17-21, Stuttgart, Germany.

(2008)

[55] McCrae, R. S., and Costa, P. T. Jr.: Adding Liebe und Arbeit: The Full Five-Factor Model and

WellBeing. Personality and Social Psychology Bulletin, 17, 227-232. (1991)

[56] Berenbaum, H., and Williams, M.: Personality and Emotional Reactivity. Journal of Research

in Personality 29, 24-34. (1995)

[57] Field, A., Hole, G.: How to Design and Report Experiments. Paperback. Sage Publications

Ltd. (2003)

[58] Ortony A., Clore G., Collins A.: The Cognitive Structure of Emotions. Cambridge University

Press, UK. (1988)

[59] Reeves, B.; Nass, C.: The Media Equation: How People Treat Computers, Television and New

Media like Real People and Places. Cambridge University Press. (1996)

[60] VanLieshout, C. F. M.: Lifespan personality development: Self-organising goal-oriented agents

and developmental outcome. International Journal of Behavioral Development, v. 24, n. 3, p. 276-

288. (2000)

[61] David, J. P., & Suls, J. (1999). Coping efforts in daily life: Role of Big Five traits and problems

appraisals. Journal of Personality, 67, 265–294.

[62] McCrae, R. R., & Costa, P. T., Jr. (1986). Personality, coping, and coping effectiveness in an

adult sample. Journal of Personality, 54, 385–405.

[63] O‟Brien, T. B., & DeLongis, A. (1996). The interactional context of problem-,emotion-, and

relationship-focused coping: The role of the Big Five personality factors. Journal of Personality, 64,

775–813.

[64] Rim, Y. (1986). Ways of coping, personality, age, sex and family structural variables.

Personality and Individual Differences, 7, 113–116.

[65] Arya, A., Jefferies, L. N., Enns, J.T., and DiPaola, S.: Facial actions as visual cues for

personality. Computer Animation and Virtual Worlds, 17: 371–382. (2006)

[66] McCrae, R. R., and Costa, P. T. Jr.:Validation of the Five-Factor Model of Personality Across

Instruments and Observers. Gerontology Research Centre, National Institute on Aging, National

Institutes of HealthJournal of Personality and Social Psychology 1987, Vol. 52, No. 1,81-90

[67] Tyler, T. R., Kramer, R. M., John, O. P.: The psychology of the social self. Lawrence Erlbaum

78

associates, publishers. (1999)

[68] Sedgwick, E. K., Frank, A., Alexander, I. E.: Shame and its sisters: a Silvan Tomkins reader.

Duke University Press. (1995)

[69] Watson, D., & Hubbard, B. Adaptational style and dispositional structure: Coping in the context

of the five-factor model. Journal of Personality, 64, 737–774. (1996).

[70] Bolger, N., & Zuckerman, A. (1995). A framework for studying personality in the stress

process. Journal of Personality and Social Psychology, 69, 890–902.

[71] Gunthert, K. C., Cohen, L. H., & Armeli, S. (1999). The role of neuroticism in daily stress and

coping. Journal of Personality and Social Psychology, 77, 1087–1100.

[72] Hooker, K., Frazier, I. D., & Monahan, D. J. (1994). Personality and coping among caregivers

of spouses with dementia. The Gerontologist, 34, 386–392.

[73] Amirkhan, J. H., Risinger, R. T., & Swickert, R. J. (1995). Extraversion: A „„hidden‟‟ personality

factor in coping? Journal of Personality, 63, 189–212.

[74] Bolger, N. (1990). Coping as a personality process: a prospective study. Journal of Personality

and Social Psychology 59, 525-537

[75] Endler, N. S., & Parker, J. D. A. (1990). Multidimensional assessment of coping: A critical

evaluation. Journal of Personality and Social Psychology, 58, 844–854.

[76] Fleishman, J. A. (1984). Personality characteristics and coping patterns. Journal of Health and

Social Behaviour, 25, 229–244.

[77] Holahan, C. J., & Moos, R. H. (1987). Personal and contextual determinants of coping

strategies. Journal of Personality and Social Psychology, 52, 946–955.

[78] Hooker, K., Frazier, I. D., & Monahan, D. J. (1994). Personality and coping among caregivers

of spouses with dementia. The Gerontologist, 34, 386–392.

[79] Lazarus, R. S., & DeLongis, A. (1983). Psychological stress and coping in ageing. American

Psychologist, 38, 245–254.

[80] Lee-Baggley, D. L., Preece, M., & DeLongis, A. (2005, October). Coping With Interpersonal

Stress: Role of Big Five Traits. Journal of Personality 73.

[81] Maitlin, J. A., Wethington, E. M., & Kesser, R. C. (1990). Situational determinants of coping

and coping effectiveness. Journal of Health and Social Behaviour, 31, 103–122.

[82] McCrae, R. R. (1990). Controlling neuroticism in the measurement of stress. Stress Medicine,

6, 237–241.

[83] Parkes, K. R. (1986). Coping in stressful episodes: The role of individual differences,

environmental factors, and situational characteristics. Journal of Personality and Social

Psychology, 51, 1277–1292.

[84] Vickers, R. R., Jr., Kolar, D. W., & Hervig, L. K. (1989). Personality correlates of coping with

military basic training (Rep. No. 89-3). San Diego: Naval Health Research Centre

[85] Vitaliano, P. P., Mairuro, R. D., Russo, J., & Becker, J. (1987). Raw versus relative scores in

79

the assessment of coping strategies. Journal of Behavioural Medicine, 10, 1–18.

[86] Wiggins, J. S., & Trapnell, P. D. (1996). A dyadic-interactional perspective on the five-factor

model. In J. S. Wiggins (Ed.), The five-factor model of personality: Theoretical perspectives (pp.

88–162). New York: Guilford.

[87] Costa, P. T., Jr., Somerfield, M. R., & McCrae, R. R. (1996). Personality and coping: A

reconceptualization. In M. Zeidner & N. S. Endler (Eds.), Handbook of coping: Theory, research,

applications (pp. 44–61). Toronto: John Wiley & Sons, Inc.

[88] Fortman, J.: The Role of Facial Expressions of Emotion in Attributions of Personality

Characteristics. Paper presented at the annual meeting of the International Communication

Association, New Orleans Sheraton, New Orleans, LA On-line. (2004)

[89] John, O. P., Donahue, E. M., & Kentle, R. L. The Big Five Inventory--Versions 4a and 54.

Berkeley, CA: University of California,Berkeley, Institute of Personality and Social Research.

(1991)

[90] Costa, P. T. Jr., McCrae, R. R.: Revised Neo-Personality Inventory (NEO-PI-R) and Neo Five-

Factor Inventory (NEO-FFI) Professional Manual. Odessa, FL: Psychological Assessment

Resources, 1992.

[91] D. Sander, D. Grandjean, and K. R. Scherer, „A systems approach to appraisal

mechanisms in emotion‟, Neural Netw., 18(4), 317–352, (2005).

[92] Ekman, P., Friesen, W. V.: Unmasking the face: A guide to recognizing emotions from facial

clues. Reprint ed. Cambridge: Malor Books. (2003)

[93] Ekman, P.: An Argument for Basic Emotions. Cognition and Emotion, v. 6, n. 3-4, p. 169-200.

Hove, UK: Lowrence Erlbaum. (1992)

[94] Clore, G. L., Storbeck, J.: Affect as information about liking, efficacy and importance. In Fogas,

J. P. (ed). Hearts and Minds: Affective influences on social cognition and behaviour New York:

Psychology Press. (2005)

[95] Eibl-Eibesfeldt, I.: Similarities and differences between cultures in expressive movements. In

R. A. Hinde (Ed.), Nonverbal communication (pp. 297- 312). Cambridge: Cambridge University

Press. (1972)

[96] Izard, C. E., Ed. Emotions in personality and psychopathology. New York: Plenum. (1979)

[97] Ekman, P.: Universal and cultural differences in facial expressions of emotion. In J Cole (Ed.),

Nebraska Symposium on Motivation: Vol. 38. Perspectives on motivation (pp. 207-283). Lincoln:

University of Nebraska Press. (1972)

[98] Keltner, D., & Ekman, P.: Affective intensity and emotional responses. Cognition & Emotion, 10

(3), p. 323-328. (1996)

[99] Gross, J., John, O.: Facets of emotional expressivity: three self-report factors and their

correlates. Person. individ. Diff. Vol. 19, No. 4, pp 555-968. (1995)

[100] LaHaye, T.: Why you act the way you do. Tyndale House Pub (1988)

80

Appendix A

Questionnaire

User test made available online includes a personality questionnaire [22], [89] and a video of a

scenario with an agent having a specific personality.

Thesis Evaluation

First of all thank you for helping us in the evaluation of our work.

We are two IST (Instituto Superior Técnico, Lisbon, Portugal) students and this is a little

questionnaire which is going to help us in the evaluation of our work.

The session should represent about 20 minutes of your time.

Our goals with this work were to build synthetic actors and give them a personality.

So, the objective of this questionnaire is to verify if the synthetic agents are or not "good actors",

and if we can see in them qualities that a good actor should have.

How I am in general

Here are a number of characteristics that may or may not apply to you. For example, do you

agree that you are someone who likes to spend time with others? Please write a number next to

81

each statement to indicate the extent to which you agree or disagree with that statement.

1 2 3 4 5

Disagree Disagree Neither agree Agree Agree

strongly a little nor disagree a little strongly

I am someone who…

1._____ Is talkative

2._____ Tends to find fault with others

3._____ Does a thorough job

4._____ Is depressed, blue

6._____ Is reserved

7._____ Is helpful and unselfish with others

8._____ Can be somewhat careless

9._____ Is relaxed, handles stress well.

11._____ Is full of energy

12._____ Starts quarrels with others

13._____ Is a reliable worker

14._____ Can be tense

16._____ Generates a lot of enthusiasm

17._____ Has a forgiving nature

18._____ Tends to be disorganized

19._____ Worries a lot

21._____ Tends to be quiet

22._____ Is generally trusting

23._____ Tends to be lazy

24._____ Is emotionally stable, not easily upset

26._____ Has an assertive personality

27._____ Can be cold and aloof

28._____ Perseveres until the task is finished

29._____ Can be moody

31._____ Is sometimes shy, inhibited

32._____ Is considerate and kind to almost

everyone

33._____ Does things efficiently

34._____ Remains calm in tense situations

36._____ Is outgoing, sociable

37._____ Is sometimes rude to others

38._____ Makes plans and follows through with

them

39._____ Gets nervous easily

42._____ Likes to cooperate with others

43._____ Is easily distracted

Now, please, watch the following video:

<video>

The following questions regard Lisa, the character of the video above. Please select a number

next to each statement to indicate the extent to which you agree or disagree with that statement.

1 2 3 4 5

Disagree Disagree Neither agree Agree Agree

strongly a little nor disagree a little strongly

In your opinion, Lisa is someone who...

1._____ Is talkative

2._____ Tends to find fault with others

3._____ Does a thorough job

4._____ Is depressed, blue

6._____ Is reserved

7._____ Is helpful and unselfish with others

8._____ Can be somewhat careless

9._____ Is relaxed, handles stress well.

11._____ Is full of energy

12._____ Starts quarrels with others

13._____ Is a reliable worker

14._____ Can be tense

16._____ Generates a lot of enthusiasm

17._____ Has a forgiving nature

18._____ Tends to be disorganized

19._____ Worries a lot

21._____ Tends to be quiet

22._____ Is generally trusting

23._____ Tends to be lazy

24._____ Is emotionally stable, not easily upset

26._____ Has an assertive personality

27._____ Can be cold and aloof

28._____ Perseveres until the task is finished

29._____ Can be moody

31._____ Is sometimes shy, inhibited

83

32._____ Is considerate and kind to almost everyone

33._____ Does things efficiently

34._____ Remains calm in tense situations

36._____ Is outgoing, sociable

37._____ Is sometimes rude to others

38._____ Makes plans and follows through with

them

39._____ Gets nervous easily

42._____ Likes to cooperate with others

43._____ Is easily distracted

Appendix B

Statistical Data

Figure B.1: Mean and deviation for the Conscientiousness trait evaluated by users

for the melancholic agent

Figure B.3: Mean and deviation for Agreeableness evaluated by users for the

melancholic agent

Figure B.2: Mean and deviation for the Extroversion trait evaluated by users for the

melancholic agent

Figure B.5: Mean and deviation for Conscientiousness evaluated by users for the

sanguine agent

Figure B.4: Mean and deviation for Neuroticism evaluated by users for the sanguine agent

Figure B.7: Mean and deviation for Agreeableness evaluated by users for the sanguine

agent

Figure B.6: Mean and deviation for Extroversion evaluated by users for the sanguine

agent

Figure B.9: Mean and deviation for Conscientiousness evaluated by users for the phlegmatic agent

Figure B.8: Mean and deviation for Neuroticism evaluated by users for the

sanguine agent

Figure B.11: Mean and deviation for Extroversion evaluated by users for the phlegmatic agent

Figure B.10: Mean and deviation for Agreeableness evaluated by users for the phlegmatic agent

Figure B.12 Mean and deviation for Neuroticism evaluated by users for the

phlegmatic agent

Figure B.13: Mean and deviation for Conscientiousness evaluated by users for

the choleric agent

Figure B.14: Mean and deviation for Extroversion evaluated by users for the

choleric agent

Figure B.15: Mean and deviation for Agreeableness evaluated by users for the

choleric agent

Figure B.16: Mean and deviation for Neuroticism evaluated by users for the

choleric agent