Upload
vanquynh
View
219
Download
0
Embed Size (px)
Citation preview
Creating Social and Cultural Agents
Samuel Francisco Mascarenhas
Dissertação para obtenção do Grau de Mestre em
Engenharia Informática e de Computadores
Júri
Presidente: Professor Doutor Nuno João Neves Mamede
Orientador: Professor Doutor Rui Filipe Fernandes Prada
Vogais: Professora Doutora Ana Maria Severino Almeida e Paiva
Professor Doutor Helder Manuel Ferreira Coelho
Maio 2009
Resumo
Com o constante aumento do desenvolvimento de agentes autonomos, surge uma crescente
procura pela capacidade dos mesmos interagirem num contexto social, de forma semelhante
aos humanos. Como resultado, varias arquitecturas de agentes comecaram a considerar diversos
factores sociais para guiar o comportamento dos seus agentes. Contudo, os aspectos culturais
tem sido bastante negligenciados, apesar de serem elementos cruciais das sociedades humanas.
Esta dissertacao propoe um modelo de cultura baseado em tres aspectos importantes de
culturas humanas: (1) Dimensoes Culturais; (2) Sımbolos e (3) Rituais. O proposito do modelo
consiste em desenvolver uma arquitectura de agentes que possibilite a criacao de diferentes
grupos culturais de personagens sinteticos. Para tal, utilizou-se uma arquitectura ja existente,
na qual se efectuou a integracao destes elementos.
Para aferir se a nova arquitectura permite expressar diferencas culturais de comportamento,
criaram-se diferentes grupos de agentes que apenas diferiram na sua cultura associada. Para
avaliar o efeito dos elementos culturais implementados na percepcao de diferentes culturas,
realizaram-se duas experiencias com os grupos criados. A primeira pretendeu avaliar o efeito
dos rituais (e sımbolos associados), enquanto que a segunda pretendeu avaliar o efeito das
dimensoes culturais. Ambas as experiencias originaram resultados positivos e significativos. Os
resultados sugerem que o modelo proposto e de facto capaz de criar culturas perceptivelmente
distintas.
Palavras-Chave: Cultura, Agentes Autonomos, Agentes Socialmente Inteligentes, Per-
sonagens Sinteticos, Ambientes Virtuais Interactivos.
Abstract
With the increasing development of autonomous agents, there is a bigger demand on their
capability of interacting with other agents in a social context, in ways that are natural and
inspired by how humans and other species interact. As a result, many agent architectures are
taking into account a plenitude of social factors to drive their agents’ behaviour. However,
culture has been largely neglected so far, even though it is a crucial aspect of human societies.
This dissertation proposes a culture model based on three important behavioural elements
of human cultures: (1) Cultural Dimensions; (2) Symbols and (3) Rituals. The purpose of
the model is to attain an agent architecture that facilitates the generation of different cultural
groups of synthetic characters. As such, we defined these elements and integrated them into
an already existent agent architecture.
With the new architecture, and in order to assess if it is possible to express different cultural
behaviour in synthetic characters, we created different groups of agents that only differed in
their associated cultures. Using these groups, two experiments were conducted to evaluate
the effect of the implemented elements in the user’s perception of different cultures. The first
experiment evaluated the effect of rituals (and associated symbols), while the second one the
effect of the cultural dimensions component. Both experiments yielded significant and positive
results, which suggest that our model is indeed capable of creating distinguishable cultures.
Keywords: culture, autonomous agents, social intelligent agents, synthetic characters,
interactive virtual environments.
Acknowledgments
First, I want to express my gratitude to everyone at GAIPS / INESC-ID for welcoming me
with such friendliness into their group. In particular, I would like to thank my supervisors,
Professora Ana Paiva and Professor Rui Prada, for presenting me with such a wonderful
opportunity to perform investigation on my favourite computer science field. Their strong
guidance, trust, and encouragement were crucial for completing this document.
Also, I would like to give a very special thanks to Joao Dias for being the one who introduced
me to GAIPS; for his contributions to this work; for constantly providing insightful comments
and suggestions; for being always available to answer any doubts I had (even late at night);
and most of all, for being such a great friend. I would also like to mention a special gratitude
to Nuno Afonso, Rui Figueiredo and Pedro Sequeira for their precious help.
Moreover, I want to thank all the members of the eCircus project team, for giving me the
opportunity to see how nice it is to work in an European project. In particular, I would like
to thank Mey Yii Lim for her contributions.
Also, I would like to thank my girlfriend, Carmen Romao. All her love, dedication, support,
and understanding made this journey much more easier.
Finally, I want to dedicate this dissertation to my family, especially to my Father who left
this world when I was still young. He will always be remembered as my personal hero. I hope
he is proud. Thank you.
i
Contents
1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Related Work 5
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Background on Culture and Synthetic Cultures . . . . . . . . . . . . . . . . . . 5
2.2.1 Geert Hofstede’s Cultural Dimensions . . . . . . . . . . . . . . . . . . . 6
2.2.2 Developmental Model of Intercultural Sensitivity . . . . . . . . . . . . . 10
2.3 Culture in Synthetic Characters . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.1 CUBE-G . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3.2 Kyra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.3 Tactical Language Training System . . . . . . . . . . . . . . . . . . . . . 15
2.4 Agent Architectures for Social and Cultural Agents . . . . . . . . . . . . . . . . 17
2.4.1 PsychSim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4.2 SGD Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4.3 CAB Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3 Conceptual Model 27
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Culture Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3 Cultural Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3.1 Goal Utility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3.2 Emotional Appraisal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
ii
CONTENTS iii
3.4 Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.5 Rituals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.6 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4 Implementation 39
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.2 FAtiMA-PSI (Baseline Architecture) . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Cultural Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3.1 Symbols Parametrisation . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.2 Symbol Translator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.3.3 Motivational State of Others . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3.4 Dimensions Parametrisation . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.3.5 Cultural Goal Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3.6 Cultural Appraisal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.7 Rituals Parametrisation . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3.8 Ritual Manager . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5 Case Studies 57
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.2 First Case Study - ORIENT Game . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.3 Second Case Study - Dinner Party . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.3.1 Character Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.3.2 Culture Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
5.3.3 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.3.4 Rituals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
5.3.5 Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.4 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6 Evaluation 67
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.2 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.3 First Experiment - Rituals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
iv CONTENTS
6.3.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.4 Second Experiment - Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . 70
6.4.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
6.5 Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
7 Conclusion 77
7.1 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
A Rituals Questionnaire 87
B Dimensions Questionnaire 91
CONTENTS v
List of Figures
2.1 The Developmental Model of Intercultural Sensitivity . . . . . . . . . . . . . . 10
2.2 Screenshot of the Mission Practice Environment . . . . . . . . . . . . . . . . . . 16
2.3 SGD Model Agent’s Architecture. . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4 Example of a socio-cultural network in CAB. . . . . . . . . . . . . . . . . . . . 22
3.1 Giving Food Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.1 FAtiMA-PSI Agent Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2 Cultural Agent Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3 Motivational States of Two Different Agents . . . . . . . . . . . . . . . . . . . . 45
5.1 Screenshot of ORIENT. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.2 Characters at the dinner table . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
5.3 Dinner Ritual (Low Power Distance Culture) - Screenshot Sequence. . . . . . . 64
5.4 Dinner Ritual (High Power Distance Culture) - Screenshot Sequence. . . . . . . 64
6.1 Results for: Do you think the differences are related to the culture or personality? 75
vi
LIST OF FIGURES vii
List of Tables
2.1 Hofstede’s ratings for ten selected countries. . . . . . . . . . . . . . . . . . . . . 9
2.2 Reviewed Systems’ Comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.1 Symbol parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.2 Cultural dimensions parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3 Ritual parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.1 First Experiment - Results for the user’s adjective classification . . . . . . . . . 70
6.2 Second experiment - Results for the statements classification. . . . . . . . . . . 72
6.3 Second experiment - Results for the user’s adjective classification. . . . . . . . . 73
viii
LIST OF TABLES ix
Chapter 1
Introduction
1.1 Motivation
Nowadays, the world is becoming a much smaller place, in the sense that people from different
cultures have more chances to contact with each other than ever before. As such, intercultural
awareness is becoming increasingly important. From a business perspective, companies are
growing globally and conducting business in countries that have very different cultures. A
cultural misunderstanding may result in a loss of an important deal, and thus companies
are starting to provide intercultural training to their employees. From a social perspective,
groups of people that have a distinct culture from the majority (usually called subcultures),
are becoming larger. Lack of understanding about these subcultures can often lead to the
creation of stereotypes that generate prejudice and discrimination.
The greater need for intercultural training has lead to the development of educating meth-
ods that are growing more demanding and complex. Passive methods such as lectures, manu-
als, films, are generally sufficient if the desired outcome is simply the acquisition of knowledge.
However, active methods are more suitable when the desired outcome is for trainees to modify
their attitudes, adopt new values and change their perspective on culture [34]. Some of these
active methods are: real-life role-plays, real-life simulations games (e.g. Barnga), or cross-
cultural dialogues. As reported in [34], each one of these methods has its own advantages and
disadvantages. For instance, real-life role-playing makes it possible to experience the role of
someone else, developing empathy and understanding, but is also very time-consuming, and in
some cases, players can be too shy to participate. Similarly, simulation games, eliminate the
gap between learning and applying, and can be used with large numbers of people. However,
1
2 CHAPTER 1. INTRODUCTION
they are also time consuming and usually require a substantial minimum number of players.
Facing these methodological issues, researchers have already begun to explore the use of
computational applications for providing new answers. In particular, there’s currently an
ongoing research on Intelligent Virtual Environments (IVEs) as another type of active method
for general educational purposes. In an IVE, similarly to a real-life role-play, a user can
experience the role of someone else while exploring a computer-generated virtual world. Also,
being a simulation game, the user gets to apply his knowledge, deciding which actions to take
given the situation. Furthermore, there is no dependency on the number of other human
participants, since the virtual world can be populated with autonomous synthetic characters
that behave in an human-like manner. Finally, the user can feel safe to face complicated social
situations and challenges, since his actions won’t have any consequences outside the virtual
world. Therefore, the significant advantage of IVEs is that they can, in theory, combine the
strengths of the active methods presented above, while simultaneously solving their related
issues.
Despite having these great advantages, the development of IVEs is still in the beginning
and there are still many challenging problems to be solved. Many of these problems come from
the fact that users want to experience the same kinds of social dynamics they would experience
in the real world, especially when interacting with synthetic characters that resemble human
beings. As such, we are witnessing a large increase on the study of agent architectures that
take into account social interactions, such as human dialogue and emotional responses.
However, culture which is a fundamental aspect of human societies, has been largely ne-
glected so far. As a consequence, the social richness of the IVEs is significantly diminished,
since the behaviour of the characters usually ends up being distinguishable only by individual
differences. While not having culture embedded in the minds of the characters can be suffi-
cient for dealing with simple real-world scenarios, we believe it is important for dealing with
more complex scenarios and essential if we wish to build IVEs that are specifically designed
for increasing intercultural awareness.
1.2 The Problem
This dissertation discusses the role of culture in the creation of groups of synthetic characters
in order to help build IVEs that are socially richer. Our objective is to address the following
1.3. OUTLINE 3
problem:
How can we build different cultural groups of autonomous synthetic characters that
exhibit distinguishable differences in their patterns of behaviour, similar to those
found in real human cultures?
To deal with this problem, we will not try to cover every aspect of human cultures due
to their enormous complexity. Instead we will focus on the more universal aspects of human
cultures that are closely related to different patterns of behaviour and reasoning. Coming
from an AI perspective, we are primarily concerned with the possibility of representing these
general aspects in an autonomous agent architecture. As such, we will not look into cultural
differences such as aesthetic styles, dialects, accents, and so forth.
Therefore, regarding the problem previously stated, our main hypothesis to be proven is:
If the behaviour of autonomous synthetic characters is driven by an agent archi-
tecture with an explicit model of culture, that is inspired by some aspects of human
cultures, users will be able to recognise cultural differences in different groups of
characters that differ exclusively in their cultural specification.
In order to prove this hypothesis we will search for anthropology studies that point out
common and relevant elements of human cultures. Based on these studies, we will define
a cultural model that encompasses such elements and apply it in an agent architecture for
synthetic characters. Afterwards, we will use the architecture to create different cultural
groups of such characters. We will then perform an evaluation to determine the impact of our
model in the user’s perception of cultural differences in the created groups.
1.3 Outline
This document is divided in six different chapters.
Chapter 2 (Related Work) starts with a brief theoretical background on culture. After-
wards, we discuss some work previously done on culture-specific synthetic characters and also
on architectures for social and cultural agents. This chapter ends with a discussion about the
systems reviewed.
Chapter 3 (Conceptual Model) discusses the conceptual model proposed for the culture
representation in the synthetic characters.
4 CHAPTER 1. INTRODUCTION
Chapter 4 (Implementation) describes an implementation of the previously defined model
extending an already existent autonomous agent architecture.
Chapter 5 (Case Studies) describes two case studies developed to evaluate the agent archi-
tecture. The first is a serious game for developing inter-cultural empathy. The second involves
the creation of different cultural groups of characters acting out a common real-life situation.
Chapter 6 (Evaluation) describes two experiments we conducted to evaluate our cultural
agents.
Finally, in Chapter 7 (Conclusion), we provide a summary of the work developed in this
dissertation and discuss some future work.
Chapter 2
Related Work
2.1 Introduction
This chapter is organised in three main sections. In the first section we will review the concept
of culture and present important anthropology studies we found relevant to achieve our goals.
In the next section, we discuss computational projects concerning the adaptation of synthetic
characters to different cultures. Then, in the last section, we review agent architectures specif-
ically designed to create agents with social and cultural behaviour. Finally, we conclude this
chapter with an analysis of the systems reviewed, stressing the influence that they had on the
development of this thesis.
2.2 Background on Culture and Synthetic Cultures
The concept of culture has been studied for many years by anthropologists and other be-
havioural scientists. In 1871, Edward B. Tylor, defined culture in [57] as ”that complex whole
which includes knowledge, belief, art, law, morals, custom, and any other capabilities and
habits acquired by man as a member of society.” Since then, many other possible definitions
have surged. In 1952, a list containing 164 possible definitions of culture was compiled by
Alfred Kroeber and Clyde Kluckhonn [33].
Nevertheless, Robert J. House et al. [27] affirms that ”despite lack of consensus among
scholars, there are several essential common threads that run throughout the various concep-
tualisations and definitions of the construct generally referred to as culture.” He considers that
culture often refers to ”collectivities in which the members share several psychological com-
5
6 CHAPTER 2. RELATED WORK
monalities - assumptions, beliefs, values, interpretations of events (meanings), social identities,
and motives - and abide by a set of shared norms in a common manner.” In other words, culture
can be broadly defined as a set of symbols and behaviour patterns that are learnt and shared
by a group of individuals. However, an important question emerges from this broad definition:
Which specific symbols and behaviour patterns establish different cultures? Anthropologists
are still debating over this issue.
Perhaps the most comprehensive and cited study about differences in cultures comes from
Geert Hofstede [24, 26]. Culture, according to him, is ”the collective programming of the mind
that distinguishes the members of one group or category of people from another” [24]. These
”mental programs” refer to patterns of thinking, feeling, and potential acting that are shared
and learnt by members of the same culture. These patterns can manifest themselves at an
implicit level, under the form of values, or at a more clearly observable level, under the form
of rituals, heroes and symbols.
The four types of cultural manifestations can be described as follows: (1) Values - represent
cultural preconceptions about what is desirable/undesirable; (2) Rituals - are essential social
activities that are carried out in a predetermined fashion; (3) Heroes - real or even imaginary
persons that serve as models for the cultural values; (4) Symbols - words, gestures, pictures,
or objects that members of a given culture have assigned a special particular meaning.
2.2.1 Geert Hofstede’s Cultural Dimensions
Asides from the cultural manifestations presented above, Hofstede proposes five dimensions on
which cultures vary [24]. Different from the previous manifestations, which can be very specific
to a certain culture or subculture (e.g. the Japanese tea ceremony), Hofstede argues that these
dimensions are universal. They are directly based on the culture’s values and indicate general
behavioural tendencies shared by the members of the culture. However, they should be not
considered deterministic, since other factors such as the individual’s personality, also play an
important role on human behaviour. We will now describe each one of these five dimensions:
1. Power Distance Index (PDI) - the degree to which less powerful members of the
group expect and accept that power is distributed unequally.
• Small PDI Cultures - in these cultures (e.g. Austria), the power relations are
usually more consultative or democratic, and people tend to regard others as equals
2.2. BACKGROUND ON CULTURE AND SYNTHETIC CULTURES 7
despite their formal status. Special privileges are usually disapproved, and powerful
people try to look the same as less powerful people. Power is usually decentralised,
and everyone’s opinion matters, including children. Moreover, people from these
cultures are usually informal and unceremonious.
• High PDI Cultures - cultures that have a high power distance (e.g. Malaysia)
tend to accept power relations that are more autocratic, and people usually respect
and acknowledge the power of others just by their formal status. Powerful people
have more privileges and like to wear symbols that reflect their status. Power is
usually centralised, and less powerful people need to ask permission to speak in a
discussion. Although very verbal, people usually talk in a soft and polite manner.
Finally, when problems arise, the tendency is to shift blame downwards in the
hierarchy.
2. Individualism (IDV) - versus its opposite, collectivism, indicates the extent to which
individuals see themselves integrated into groups.
• Collectivistic Cultures - people from these cultures (e.g. Guatemala), are usually
integrated into strong and cohesive groups that are an integral part of their identity.
In these groups, everyone looks out for one another in exchange for unquestioning
loyalty. Therefore, relationships are very important, and the harmony of the group
should always be maintained, avoiding direct confrontations. Moreover, laws and
right are expected to differ by group.
• Individualistic Cultures - these cultures (e.g. USA), tend to stress the impor-
tance of personal achievements and individual rights. People are expected to be
only responsible for themselves and their immediate family. Self-definition is based
on personal cues (personality, values, and physical attributes). Relationships are
less important than the task at hand, and other people tend to be measured by
their usefulness. Moreover, when in groups, people like to stand out visually.
3. Masculinity (MAS) - versus its opposite, femininity, refers to the distribution of roles
between genders.
• Feminine Cultures - in a very feminine culture (e.g. Sweden), relationships and
quality of life are very important. Men and women are supposed to be modest,
8 CHAPTER 2. RELATED WORK
soft-spoken, and care for the weak. Compromise and negotiation are used to solve
conflicts. Both sexes should have equal rights and responsibilities. Moreover, people
are generally warm and friendly in a conversation.
• Masculine Cultures - a very masculine culture (e.g. Japan) favours assertiveness,
ambition, efficiency, competition and materialism. Also, differences between gender
roles are accentuated. Men are supposed to be ambitious and tough, while women
should be subservient and tender. Conflicts are settled by arguing or fighting them
out. Moreover, people are generally hard to please, tend to be overachievers, and
blame others for their mistakes.
4. Uncertainty Avoidance Index (UAI) - this dimension indicates to what extent
people prefer structured over unstructured situations.
• Low UAI Cultures - cultures with a low UAI (e.g. Singapore), have as few rules
as possible, and people are more tolerant of opinions different from what they are
used to. Unfamiliar risks and ambiguous situations cause no discomfort, and people
have few taboos. Moreover, people are generally patient, relaxed, informal, and tend
to work hard only when it is needed. Open-ended questions and innovative ideas
are very common in conversations. Finally, emotions and aggression are usually
hidden.
• High UAI Cultures - these cultures (e.g. Portugal), tend to have strict laws and
rules, and also various safety and security measures to avoid situations that are
novel, unknown, ambiguous, surprising, or different from usual. At an emotional
level, people are more expressive and motivated by their inner nervous energy. Usu-
ally, these cultures have rigid beliefs and strong taboos. Moreover, people tend to
be organised and seek specialisation.
5. Long-Term Orientation (LTO) - indicates to what extent the future has more im-
portance than the past or present.
• Short-Term Oriented Cultures - respect for tradition, quick results, fulfilling
social obligations and reciprocation of gifts and favors are greatly valued in short
term oriented cultures (e.g. Nigeria). People are ceremonious, live day by day, and
usually talk a lot, particularly about the past.
2.2. BACKGROUND ON CULTURE AND SYNTHETIC CULTURES 9
• Long-Term Oriented Cultures - in these cultures (e.g. China), people give
more importance to the future than the past and present. Pragmatism, thrift, hard
working, prosperity and perseverance are greatly valued. People may devote their
lives to philosophical ideals. Moreover, people tend to talk in a direct and focused
manner.
The foundation for Hofstede’s theory is a large empirical study of IBM’s employees in more
than 70 countries analysed. Table 2.1 shows the ratings (the * indicates that the value is
unknown) of the ten countries that were used above as examples.
PDI IDV MAS UAI LTO
Austria 11 55 79 70 *
China 80 20 66 40 118
Guatemala 95 6 37 101 *
Japan 54 46 95 92 80
Malaysia 104 26 50 36 *
Nigeria 77 20 46 54 16
Portugal 63 27 31 104 *
Singapore 74 20 48 8 48
Sweden 31 71 5 29 33
US 40 91 62 46 29
Table 2.1: Hofstede’s ratings for ten selected countries.
In Exploring Culture [26], Hofstede et. al. introduced ten synthetic cultures, which are ex-
treme manifestations of the value orientations at both ends of the cultural dimensions presented
above. Real cultures, unlike the synthetic ones, have elements of all dimensions, and may not
fall in the extreme side of any particular dimension. Synthetic cultures, on the other hand,
simplify the complex notion of culture, by isolating the behavioural tendencies specific to each
extreme. Therefore, they can be used as an intercultural training technique to simulate cross-
cultural encounters. Moreover, since synthetic cultures aren’t literal representations of any
culture from the real world, it is more unlikely that trainees will object to the generalisations
made.
10 CHAPTER 2. RELATED WORK
2.2.2 Developmental Model of Intercultural Sensitivity
To study the effectiveness of an intercultural training tool, it is essential to have a good
understanding of the different predispositions and preconceptions a person has on cultural
differences. Some techniques may present excellent results in a person that already accepts the
existence of these differences and fail miserably if such existence is denied. In order to achieve
this understanding, the developmental model of intercultural sensitivity (DMIS), which was
built by Milton Bennett [8, 9, 10, 34], provides a valuable framework to explain people’s
different reactions in the presence of intercultural situations. He examined several students
over the course of months and sometimes years in various intercultural workshops, classes,
exchanges, and graduate programs. In this study, he detected common behavioural tendencies
among the students, and based on grounded theory, he organised these tendencies into six
stages of increasing sensitivity to cultural difference (see Figure 2.1).
Figure 2.1: The Developmental Model of Intercultural Sensitivity
The first three DMIS stages are ethnocentric, meaning that one’s own culture is experienced
as central to reality in some way:
• Denial - people at this stage will generally show a lack of interest in cultural differences,
experiencing their own culture as the only real one. Reasoning about other cultures is
avoided by maintaining psychological and/or physical isolation and attempts to address
intercultural issues head-on are likely to result in bewilderment and even hostility.
• Defense - one’s worldview is divided into ”us and them” and one’s own culture is
2.2. BACKGROUND ON CULTURE AND SYNTHETIC CULTURES 11
experienced as being the only good one. People tend to demonstrate an ”under siege”
attitude by heavily criticising other cultures. From an outsider point of view, members
of the host culture seem to be trying to defend their supposed cultural privileges even
though they may not think that way at a conscious level. On the other hand, members
of the guest culture feel as if their cultural identities are being threatened by the pressure
to assimilate their host’s culture.
• Minimization - a universal view of culture is adopted and cultural divergences are
trivialised to a superficial level. Differences in etiquette and customs are recognised but
more complex aspects of one’s behaviour are thought to be similar throughout all cultures.
Since others are viewed as similar, people at the Minimization stage try excessively to
correct the behaviours of other culture’s members.
The second three DMIS stages are ethnorelative, meaning that one’s own culture is expe-
rienced in the context of other cultures:
• Acceptance - at this stage, other cultures are experienced as equally complex but
having different constructions of reality. People, in general, will demonstrate curiosity
and respect for cultural differences. However, the acceptance of these differences does
not translate directly to their agreement or liking.
• Adaptation - people at this stage are capable to take the perspective of another cul-
ture worldview, thus they may deliberately adapt their behaviour to communicate more
effectively in an intercultural situation. This ability is seen as cultural empathy, which
can be described as ”the attempt to organise experience through a set of constructs that
are more characteristic of another culture than of one’s own.” [34] A more elaborate
discussion about this type of empathy is found in [10].
• Integration - the final stage of the DMIS is commonly found in members of minority
cultural groups or long-term expatriates. Even though people at this stage are multicul-
tural in their worldview, they aren’t necessarily better in intercultural situations, since
they might be unable to select appropriate behaviour in a given cultural setting. People
in this stage often need to reestablish their identity, which was somewhat lost in their
cultural multiplicity in order to embrace their extended experience of culture.
12 CHAPTER 2. RELATED WORK
The fundamental assumption of the model is that intercultural competence increases as
one’s experience of cultural difference becomes more complex and sophisticated. As we have
seen, each stage represents a different development of the cognitive structure, reflected by
different kinds of attitudes. While the ethnocentric stages are related to avoidance of cultural
difference, the ethnorelative ones can be seen as ways of seeking cultural difference.
The theoretical basis of the DMIS is personal construct theory and its extension, radical
constructivism. Personal construct theory was formulated by George Kelly, who states that
experience is a function of our classification, or interpretation of events. According to this
theory: ”A person can be a witness to a tremendous parade of episodes and yet, if he fails
to keep making something out of them, he gains little in the way of experience from having
been around when they happened. It is not what happens around him that makes a man
experienced; it is the successive construing and reconstruing of what happens, as it happens,
that enrich the experience of his life.” [32]
This model has been successfully employed in several intercultural education and training
programs. Trainers use it to diagnose client’s readiness to different types of training and to
choose and sequence appropriate training strategies. For our work, the model can be useful
to measure one’s development of intercultural sensitivity, when interacting with cultural and
social agents.
2.3 Culture in Synthetic Characters
A substantial part of the work done on culture in synthetic characters involves the adaptation
of the characters to a particular culture or to the user’s culture. In this line of work, we
decided to review GUBE-G, a project that uses Hofstede’s dimensions for inferring the cultural
background of a user (using Nintendo’s Wii remote controller), and for adapting the nonverbal
behaviour of embodied conversational agents according to the user’s culture. Afterwards, we
look at the framework used to design Kyra, a synthetic character adapted for three different
countries, discussing cultural differences that were taken into consideration for the character’s
design. Finally, we discuss the Tactical Language Training System, an IVE where users interact
with characters from a foreign culture in order to train the culture’s spoken language and
gestures.
2.3. CULTURE IN SYNTHETIC CHARACTERS 13
2.3.1 CUBE-G
CUBE-G [50] stands for ”CUlture-adaptive BEhavior Generation for interactions with embod-
ied conversational agents”. It is an interdisciplinary project funded by the German Research
Foundation (DFG). Even though CUBE-G focus exclusively in expressive behaviour, it is one
of the few known computational systems that parameterise cultural behaviour in synthetic
characters. The main goal of the project is to build a system where embodied conversational
agents are capable of adjusting their expressive behaviour to the user’s culture. To achieve this
goal, two tasks are addressed: (1) defining the cultural specific behaviour, and (2) inferring
the user’s cultural background trough the use of sensors.
For the first task, the computation model of culture used is based on the cultural dimensions
of Hofstede [9], and the correlations suggested by him in [10], between extreme positions on
each dimension and the following six variables of expressive behaviour: (1) overall activation
- number of gestures in a specific time; (2) spatial extent - how much space a gestures uses;
(3) speed - temporal extent of movements; (4) power - the strength of gestures; (5) sound -
how loud the agents speak; (6) distance - how far apart agents stand while they interact. Note
that the first four variables were used before by Bevacqua [11] in defining the expressivity of
gestures.
For the last task, due to the functionality of the sensor used (Nintendo’s Wii remote
controller), only the gestural expressivity (overall activation, spatial extent, and speed) of the
user is analysed. Gaze and speech, although also considered important aspects of observable
behaviour for deriving the user’s culture, are left out of the analysis.
Since there might be a specific user that deviates from his cultural prototypical behaviour,
and given the imperfections of the recognition system used, the model has to deal with unre-
liable and incomplete information. This led to the decision of using Bayesian networks which,
as described in [28] , are a formalism to represent probabilistic causal interactions.
To better understand how the model works, we will give the example presented in [50]: Let’s
suppose that the user’s gestures are slow, not powerful, not extended in space, and the overall
activation is unknown (only one gesture was detected). With these evidences, the Bayesian
network is updated to allow for inferring, in a diagnostic way, the user’s cultural Background.
The probability is then propagated via the dimensional nodes to the culture node, and the
system estimates the user to belong to a Swedish culture (see Table 2.1 for Sweden’s ratings).
14 CHAPTER 2. RELATED WORK
Afterwards, the agent’s behaviour is set accordingly: they stand far away from each other,
speak in a mid voice, do not gesture much, and when they actually do a gesture, they do it
slowly and with little spatial extent. If the user’s gestures were slower and wider, than the
inferred culture would be Chinese and the agents would move closer, and would use more,
wider and more powerful gestures.
2.3.2 Kyra
Kyra [39] is a synthetic character with autonomous behaviour and personality traits, devel-
oped by students and researchers at Stanford University’s School of Education and Computer
Science Department and Extempo Systems Inc. Kyra is presented as trendy girl with about
twelve years old, which attempts to motivate and educate preteens, about artistic expression
values and art history tendencies. Users interact with the character through a website where
they can communicate by typing text messages. Kyra, on the other side, communicates by
gestures, textual, and spoken utterances. Furthermore, Kyra has a complex mechanism for
understanding natural language, a mood system, and a user’s model that allows her to respond
in an appropriate manner to the user’s sentences.
The main goal in Kyra’s design was to create a believable character that could give the
illusion of life through the user’s suspension of disbelief [56] (the user perceives the character
as if it has a life of his own, dismissing its artificial nature). Initially, the character was built
for an American audience, but later a Venezuelan and Brazilian Kyra were created, aiming to
better appease the users from those countries.
To achieve believability, the authors propose a framework of ten key characteristic qualities:
(1) identity, (2) backstory, (3) appearance, (4) content of speech, (5) manner of speaking, (6)
manner of gesturing, (7) emotional dynamics, (8) social interaction patterns, (9) role and (10)
role dynamics. These characteristics ”both define and are defined by each character’s unique
idiosyncratic behaviors and signature personality traits, as well as by the character’s cultural
grounding.” [39]
Interestingly, to maintain the character’s believability in its cultural adaptation, all the ten
qualities required changes. We will now discuss the cultural implications found by the authors,
related to only two of these qualities (the ones we found more related to our goals):
1. Manner of Gesturing - indicates how the character expresses itself nonverbally. Ges-
2.3. CULTURE IN SYNTHETIC CHARACTERS 15
tures are common in all cultures, and are considered to be a fundamental part of human
dialogue. Still, there are substantial cultural differences related to gestures, such as their
diversity and frequency of use. Furthermore, identical gestures often have different or
even opposite meanings in different cultures. For example, in Bulgaria, nodding one’s
head means ”no”, while shaking it means ”yes”. Likewise, in Japan or Korea, maintain-
ing direct eye contact in a conversation can be insulting, whereas, in the United States,
avoiding it can be a sign of dishonest behaviour or shyness.
2. Emotional Dynamics - encompasses the character’s emotional state, which is subse-
quently expressed in the character’s behaviour. The basic emotions theory, grounded on
primate and cross-cultural studies, indicates that emotions such as fear, anger, sadness,
joy, disgust, and maybe surprise and interest are shared and acknowledged by all humans
[21]. Nevertheless, factors such as the appropriateness, frequency, or length of time an
emotional state lasts, vary across cultures and were taken into account in Kyra’s adap-
tation. For instance, Venezuelan Kyra maintains every emotional charge for the longest
time among the three.
2.3.3 Tactical Language Training System
In the Introduction chapter, we have discussed the advantages of using IVEs as an educational
tool. We will now analyse such a system for intercultural training on communicative skills
that illustrates the strength of these advantages. The Tactical Language Training System
(TLTS) [29, 30] is an IVE developed at the University of Southern California, USA. The goal
is to teach communicative skills in foreign languages that are less commonly taught, such as
Arabic, Chinese or Russian. Learning such languages with traditional courses can be very
time-consuming, due to their unfamiliar writing systems and cultural norms. This can partly
explain why less than 1% of USA college students enrol in Arabic courses [44].
Being presented as a serious game, the TLTS is a very different approach from a traditional
course. It has two main components: a Mission Skill Builder (MSB), and a Mission Practice
Environment (MPE). The first one consists mainly in a set of interactive exercises, in which
learners practice saying common words and phrases into a microphone, which is then auto-
matically analysed for pronunciation errors. A virtual tutor is also present to give immediate
feedback and useful suggestions.
16 CHAPTER 2. RELATED WORK
Figure 2.2: Screenshot of the Mission Practice Environment
In the MPE (Figure 2.2), learners test out their communicative skills by assuming the role
of an Army Special Forces unit character and exploring a virtual village, where they must
complete a series of missions. The missions require the player to build rapport among the
natives. These natives are autonomous synthetic characters with their own agendas. They will
only collaborate with the player if they trust him. To gain their trust, the player must speak
correctly to them and use proper cultural gestures. There is also a special aide character that
can help the user by suggesting him what to do. Figure 2.2 exemplifies a nonverbal interaction
where the player is greeting a local with a proper gesture to then ask for the village’s leader
whereabouts (a mission objective).
The first release of the system was called Tactical Iraqi (adapted to the Iraqi Arabic dialect).
An evaluation performed by the U.S. Special Operations Command found that trainees learnt
this type of Arabic to a proficiency level of a novice-high according to the American Council
on the Teaching of Foreign Languages. Many users rated the system better and more fun than
other self-study and instructor-led classes. Due to its success, adaptations to other languages
were developed, such as Tactical Pashto and Tactical Dari.
However, regarding our goals, the most relevant aspect of the TLTS is the architecture
that controls the behaviour of the virtual characters. That architecture is called Thespian [54]
and it was built on top of PsychSim [49], a general agent framework capable of generating
2.4. AGENT ARCHITECTURES FOR SOCIAL AND CULTURAL AGENTS 17
social and goal-oriented behaviour (we will review this framework in greater detail in the next
section). Interestingly, the way Thespian embeds cultural norms in the characters’ conduct
is by using social relationships such as trust and liking and then by authoring special social
variables to represent temporary obligations between agents. Obligations are created when
an agent perform a certain action on another agent. To satisfy the obligation the target
agent must choose a proper action in response. Examples of such obligations are: greeting
and greeting back, thanking and saying you are welcome, offering and accepting/rejecting etc.
These obligations somewhat resemble the notion of cultural rituals proposed by Hofstede. By
giving goals to fulfil these obligations, agents will likely follow the encoded cultural norms and
expect other agents (including the user) to follow them as well.
2.4 Agent Architectures for Social and Cultural Agents
Research on agent architectures that include social and cultural factors in their agents’ internal
knowledge and reasoning is quite new. Some of the existent computational models of culture
only study the evolution of culture using simple multi-agent models (like the one proposed
by Axelrod in [3]). However, they do not address the ways culture affects the behaviour of
individuals, which is our main focus.
In the Tactical Language Training System analysis, we’ve seen that the behaviour of the
characters was greatly driven by PsychSim, a architecture for social behaviour. In fact, culture
is fundamentally a social construct. With that in mind, we decided to analyse the characteris-
tics of PsychSim that made it a good choice to extend it with cultural obligations in Thespian.
We will also look to another social model called SGD that explores the notions of social power
and context. Finally, we will examine the CAB model, an agent architecture that mainly aims
to provide a framework where ethnographers can encode explicit cultural norms in the agents’
reasoning and knowledge.
2.4.1 PsychSim
PsychSim [49] is a multi-agent system in which the user is capable of defining a social scenario
to explore how a diverse set of entities (groups or individuals) interact, and to see how those
interactions can be influenced. Each of these entities has the following properties: goals,
preferences, relationships (e.g. friendship, hostility, and authority), private beliefs and mental
18 CHAPTER 2. RELATED WORK
models about other entities. The simulation tool generates the behaviour for these entities
and explains the outcome in terms of entity’s preferences and beliefs. The user can introduce
variations on the scenario and specify actions or messages for any entity to perform. Therefore,
PsychSim allows the user to explore multiple tactics for dealing with a social issue and to see
potential consequences of those tactics. School bullying is one particular social issue in which
PsychSim has been tested. Also, as we have said earlier, PsychSim is responsible to drive the
behaviour of the virtual characters in the Tactical Language Learning System.
PhsychSim implements a social theory called Theory of Mind, which is defined in [45] as the
human ability of attributing mental states such as intentions, beliefs, and values, not only to
oneself but to others as well. These mental models of others can greatly affect human behaviour,
and in the absence of accurate information, humans use stereotypes to form these models. In
PsychSim, agents have also a mental model of other agents, representing a subjective view of
the other agent’s goals, beliefs and policies. Moreover, these mental models are initialised with
pre-defined prototypes.
For the bullying scenario, some of the prototypes created correspond to: selfishness, altru-
ism, dominance-seeking. As an example, if Agent A thinks that Agent B is selfish, then Agent
A infers that self-wealth is very important for Agent B. On the other hand, if the Agent B is
perceived as altruistic, then Agent A infers that Agent B has the goal of helping the weak.
Agents are capable of influencing another agent’s beliefs by exchanging messages. When
an agent receives any messages, he considers whether to accept them and make the necessary
change in his beliefs, or to reject them. This decision is based on the following three factors:
• Consistency - the agent evaluate the degree to which a potential belief conforms to
prior observations by asking itself, ”If this belief holds, would it explain the past better
than my current beliefs?” To know the answer, the agent assesses the quality of the
competing explanations by a re-simulation of the past history.
• Self-interest - two sets of beliefs are compared, one in which the message is accepted
and one in which is rejected. The agent will favour the set that will bring him closer to his
goals in the future. The speaker’s self-interest is also evaluated. If the speaker benefits
greatly that the recipient believes in the message sent, then the chances of rejection by
the recipient will increase.
• Bias - corresponds to the level of trust and support that exists between the speaker and
2.4. AGENT ARCHITECTURES FOR SOCIAL AND CULTURAL AGENTS 19
recipient. It serves as a tie-breaker when consistency and self-interest both fail to decide
whether to accept the message or reject it. Trust in another agent is increased/decreased
whenever a message from that agent is accepted/rejected. Similarly, an agent increases
its support level for another agent every time the second agent selects an action that as
a high reward, concerning the first agent’s goals.
PsychSim’s agents are capable of interacting in a social manner by forming a mental model
of other agent’s beliefs and by being able to influence those beliefs. These ideas can prove to
be useful in designing our model.
2.4.2 SGD Model
The Synthetic Group Dynamics Model (SGD Model) [47, 48] is a model designed to characterize
and drive, in a believable way, group interactions in social groups formed with autonomous
agents. Figure 2.3 shows a diagram of the agent’s architecture.
Figure 2.3: SGD Model Agent’s Architecture.
The model focuses on small groups, without a strong organizational structure, that are
committed to the resolution of collaborative tasks. It is based on theories of groups dynam-
ics, developed on human social psychological studies, in particular the ones developed by
Cartwright and Zander [12], Bales [5] and McGrath [41]. The knowledge that agents possess,
in order to implement the SGD Model in their behaviour, is divided into four distinct levels:
• The individual level - defines the individual characteristics of each group member,
such as their abilities and personality. The agent’s abilities determine the actions that
20 CHAPTER 2. RELATED WORK
each agent is capable of performing, as well as the corresponding level of expertise that
the agent has in performing each of these actions. The agent’s personality is defined
using two of the dimensions proposed in the Five Factor Model [19]: (1) Extroversion -
represents the dominant initiative of the agent, and influences the agent’s frequency of
interaction; (2) Agreeableness - defines if the agent will favour positive/negative socio-
emotional interactions, depending if the agent has a high/low value in this dimension.
• The group level - defines the group members and their attitude towards the group,
as well as the underlying structure of the group, which is defined in two dimensions:
(1) the structure of power - this dimension emerges from the social influence relations
which determine the power agents have to influence the behaviour of another and (2) the
structure of interpersonal attraction - this dimension emerges from the social attraction
relations (likes/dislikes) that exist between each group member. Each member has a
relative position in the group that indicates his significance in the group. The agents
that have a higher position in the group will usually be the targets of positive socio-
emotional interactions, while the targets of negative socio-emotional interactions will be
the agents with a lower position. This position depends on: the overall social influence
that an agent has on others; the attraction felt by other agents towards him; and his
relative level of expertise. Furthermore, agents also build a relation with the group itself,
having a level of motivation for the group’s interactions and a level of attachment to the
group. Finally, the group has also a unique name to allow the agents to recognize it and
refer to it.
• The interaction level - describes the knowledge that the agents build concerning the
interactions of the group, which occur when agents execute actions that can be perceived
and evaluated by the group’s members. The interactions are responsible for creating
the dynamics in the group and the frequency of them depends on the agent’s motiva-
tion, group position and personality. They can be separated in two main categories: (1)
socio-emotional interactions that are linked to the social relations or (2) instrumental
interactions that are connected to the task the group is trying to complete. Both cate-
gories are classified as positive/negative if they provoke a positive/negative reaction by
the agents. The reactions from the agents are translated to changes in the structures of
the group (instrumental interactions affect the structure of power and socio-emotional
2.4. AGENT ARCHITECTURES FOR SOCIAL AND CULTURAL AGENTS 21
reactions affect the structure of interpersonal attraction).
• The context level - describes the task model and the knowledge about the agent’s envi-
ronment. Additionally, social norms that influence the interpretation (positive/negative)
of the social-emotional interactions may also be defined in the context level. However,
the model does not have any mechanism for the creation of these social norms, so they
have to be defined for each specific context.
The SGD Model provides an interesting way for creating believable group dynamics on
groups of synthetic characters, implemented by autonomous agents. In order to achieve this,
agents build social relations of power and interpersonal attraction with each other. Moreover,
agents have the notion of belonging to a group and are attributed a position in it, according to
their relative importance. The more important agents are the ones with a higher social status
and/or level of expertise. The model also includes a cultural aspect in the behaviour of the
agents by allowing to define social norms and rules in the context level. However, these rules
and norms are equally adopted by all agents and they do not change dynamically.
2.4.3 CAB Model
The Culturally Affected Behaviour (CAB) model [55] is a recent agent model of human be-
haviour that explicitly represents socio-cultural knowledge and reasoning. The aim of the
model is to be able to encode ethnographic data on cultural norms, biases and stereotypes,
which can then be used to drive the behaviour of synthetic characters. The model is grounded
on two social theories: Theory of Mind [45] and the Schema Theory [14].
We have already seen the use and importance of the Theory Of Mind in the PsychSim
system. In CAB, this theory was also necessary, in particular to model explicit cultural stereo-
types and biases. As for modelling cultural norms, the authors were inspired by the Schema
Theory proposed by D’Andrade. This theory postulates that a culture can be represented as a
shared organisation of schemas. Schemas are an old concept, introduced by Kant in [31]. Many
other terms have been used with similar meanings, such as ”frames” in [43] or ”scripts” in [53].
The basic idea is that a schema associates an abstract concept to a collection of knowledge
around it. For instance, the ”writing” schema is associated with someone using an pointed
object that leaves a trace across a surface (the object and the surface are left undetermined).
To describe cultural norms, D’Andrade presents in [13] a type of cognitive schema: the
22 CHAPTER 2. RELATED WORK
constitutive rules system. This schema defines a set of rules that are known, shared, and
adhered to by members of a culture. When someone performs an action that is related to any
such rule, the corresponding schema is triggered by the members of the culture.
CAB proposes a way of modelling cultural norms by representing constitutive systems of
rules through socio-cultural networks. Figure illustrates an example of such a network for the
Iraqi Sunni culture.
Figure 2.4: Example of a socio-cultural network in CAB.
The rectangular nodes represent tasks that the agent may perform in the environment and
the rounded nodes represent states that are associated to norms. The intrinsic utility values of
the states here represents the shared importance that the members of the culture place on the
associated norm (the higher the number the more important the norm is). A state can also
have a negative intrinsic utility, meaning that the associated norm has a negative connotation
in the culture. The lines represent the effect a certain task has on a state. Depending on their
sign (+/-), effects determine if the execution of a task reinforces/diminishes the current utility
of the state, by a certain degree. The current utility determines to what extent the norm is
currently being satisfied or not. Also, note the states can refer to the agent’s perspective of
himself or the agent’s perspective of others.
The authors propose that the current utility of the network, called the Socio-Cultural
Satisfaction (SCS), can be used to influence the behaviour of an agent in different ways (the
value is calculated by summing the utility of all the states in the network). For example,
imagine an agent that is a synthetic character in a virtual meeting with a human user. The
2.5. CONCLUDING REMARKS 23
agent can respond more favourably with the user if the SCS has a high value and vice versa.
Also, the agent can prefer plans that does not involve collaboration with the user in the case
of a low SCS. Finally, the character can decide to end the meeting prematurely if the SCS falls
below a certain threshold.
In relation to our work, the CAB model has some similarities. Namely, it is based on
studies from social sciences on real human cultures and also aims to provide an explicit model
of culture to change the behaviour of autonomous synthetic characters. However, the focus of
CAB is to allow a ethnographer the encoding of very specific cultural norms, which are tied to
very specific actions (such as giving alcohol, talking about cars, etc). Instead, our work aims
to provide a more general model of culture that is not focused in such specific norms.
2.5 Concluding Remarks
Our goal is to build an agent architecture that is capable of modelling social groups of agents
that have specifiable cultures, with different ways of behaving associated. However, culture
is a vast concept that is not easy definable, thus we started by reviewing a cultural model:
Hofstede’s Cultural Dimensions. The main advantages of this model are: it gives a clear and
detailed notion of the differences between national cultures; it is based on a large empirical
survey; and it is relatively easy to use in a computational way.
However, Hofstede’s model focuses exclusively in cultural differences derived from the in-
dividual’s nationality, and it has been criticised by some anthropologists for its accuracy. The
main reason for the majority of these criticisms is that the model assumes that within each
nation there is a uniform national culture, shared by the entire population, and it also assumes
that a small group of IBM’s employees are representative of that culture. Thus, it disregards
the fact that the national cultural characteristics of an individual can be influenced and modi-
fied by their membership in ethnic, religious, and social groups that have their own specialised
cultures. Despite the criticisms, we conclude that the model serves the purposes of our work,
since we want to characterise cultural behaviour and not to replicate national cultures in the
most exact way.
We also reviewed the DMIS model, which offers a way to understand people’s different
predispositions and reactions to cultural differences, by organising them into six different stages
of increasing sensitivity to cultural distinctions. Considering our goal, the model can provide an
24 CHAPTER 2. RELATED WORK
approach to evaluate the efficacy and appropriateness, of an Intelligent Virtual Environment,
as an intercultural training tool. However, this evaluation is somewhat subjective, since the
model does not clearly specify how to perform it.
Afterwards, we analysed some computational systems that focus on cultural and social
issues. We started with CUBE-G, Kyra, and TLTS, which are systems where synthetic char-
acters are adapted to different cultures. Then we reviewed PsychSim, the SGD Model, and
the CAB model. These are agent architectures that focus on social and cultural behaviour.
Table 2.2 synthesises the fundamental aspects of all these systems.
System
Cultural
Nonverbal
Behaviour
Cultural
Verbal
Behaviour
Cultural
Goal-
driven
Behaviour
Culture
Parame-
terization
Cultural
Emotional
Behaviour
Social
Relations
and Inter-
actions
Group
Structure
CUBE-G Yes No No Yes No No No
Kyra Yes Yes No No Yes No No
TLTS Yes Yes Yes Yes No Yes No
PsychSim No No No No No Yes Yes
SGD No No No Yes No Yes Yes
CAB Yes Yes Yes Yes No No No
Table 2.2: Reviewed Systems’ Comparison.
CUBE-G uses Hofstede’s Dimensions to parametrise culture in order to dynamically adapt
the nonverbal behaviour of embodied conversational agents to the user’s cultural background.
The system provides a good insight on how to correlate the cultural dimensions to specific
expressive behaviour. However, the model focuses exclusively on this type of behaviour.
Kyra, on the other hand, was specifically designed for three different cultures: American,
Brazilian, and Venezuelan. Asides from nonverbal behaviour, the authors present nine other
character qualities, where cultural distinctions are present and were taken into consideration
for the character’s believability. Unfortunately, the authors do not provide an easy method to
adapt these qualities to different cultures.
While we think the cultural adaptation of synthetic agents to the user’s culture is a very
challenging topic and CUBE-G and Kyra gives us quite helpful insights into ours goals, the
main objective of our work is not to build a system that adapts to the user’s culture but rather
to achieve an agent architecture that can easily create different cultures whenever used for
different cultural contexts.
The Tactical Language Training System is more related to our purposes, since its synthetic
characters are authored with a specific culture in mind. However, since its main purpose
2.5. CONCLUDING REMARKS 25
is language training, it only addresses communicative aspects of a culture, namely spoken
language and gestures. Interestingly, characters have goals to fulfil cultural obligations, a
notion we believe to resemble rituals.
PsychSim, the base architecture used in TLTS creates agents with the Theory of Mind
ability. We believe that this ability is also useful to our work, in particular to model collec-
tivistic cultures, where characters care about the consequences their actions have on other
group members.
Then, we find the SGD Model to be interesting due to its notion of group, and the social
relations of power and interpersonal attraction that exist between the group’s members. As
Hofstede suggests in his cultural model, these types of social relations are greatly affected by
culture. While the model does offer the possibility to define some cultural parameters, this
parametrisation is very limited for our purposes.
Regarding the Culturally Affected Behaviour model, while it is an agent architecture that
has an explicit model of cultural behaviour, the model is mainly focused on the definition of
specific cultural norms. Instead, we wish to explore more general aspects of cultural behaviour
that are not linked to particular tasks or actions, such as Hofstede’s cultural dimensions. Also,
note that CAB’s socio-cultural networks indicate only the agent’s cultural satisfaction level.
The direct consequences this level has on the agent’s behaviour are not formally specified.
Finally, after analysing all these systems, we conclude that they have important features
for our work, but none of them addresses all our needs.
26 CHAPTER 2. RELATED WORK
Chapter 3
Conceptual Model
3.1 Introduction
This chapter presents a conceptual model to explicitly model synthetic cultures in groups of
autonomous characters that are driven by an agent architecture. For simplicity reasons, the
term character is used to refer to an agent as well. To explain our conceptual model, we will
define the elements that specify a culture and describe their respective implications on the
behaviours of the characters.
3.2 Culture Specification
The main inspiration for our model comes from Hofstede’s ideas on culture, in particular in
his dimensional model [24] which we have discussed previously. As such, one of its fundamen-
tal principles is that behavioural tendencies are shared by the same characters with the same
culture, and those tendencies are dependent on how the culture is rated on a set of dimensions.
Therefore, we will define the use of these dimensions to create synthetic cultures that share
similar behavioural tendencies to those found in human cultures. Furthermore, aside the cul-
tural dimensions, our model also encompasses two other cultural manifestations identified by
Hofstede: (1) Rituals and (2) Symbols. As stated earlier, Hofstede argues that these mani-
festations are more clearly observable than the ones originated by the dimensions (people are
generally more aware of them). As such, they are included to have a richer model of culture,
but also to compare them against cultural dimensions in the user’s identification of different
cultures.
27
28 CHAPTER 3. CONCEPTUAL MODEL
Hence, a culture c in our model is defined as a 3-tuple 〈D,S,R〉 where:
• D contains the dimensional scores for Hofstede’s dimensions.
• S specifies a set of symbols that have a cultural meaning associated.
• R correspond to the set of rituals that are performed in the culture.
Notice that we will not consider Hofstede’s notion of Heroes, since we do not think it is
fundamental for our problem, besides being a very complex notion to successfully model. Fur-
thermore, we do not explicitly model cultural Values. However, they are still implicitly mod-
elled since the cultural dimensions are directly based on them. For instance, self-independence
is a value related to a high individualism dimension score. As such, the behavioural tendencies
associated to a highly individualistic culture emphasise that value.
3.3 Cultural Dimensions
Hofstede’s model has five different cultural dimensions which normally range from 0 to 100.
Our intention is to use these values to change the agent’s behaviour in a way that is congruent
with Hofstede’s findings. As described in the Related Work chapter, CUBE-G already maps
these dimensions to expressive nonverbal behaviour. We wish to pursue a different approach.
As such, we decided to use the dimensions to influence two other important aspects that are
usually present in an autonomous character: (1) goal utility and (2) emotional appraisal. The
first one is used for a character to make more rational decisions about what he should do at any
given moment. The latter serves to simulate human emotional responses to events, which is
also a fundamental requirement for the believability of the character (e.g. in Kyra, emotional
dynamics are defined as one of the key characteristic qualities to achieve believability). How-
ever, for simplification purposes, we decided to encompass only two of the five dimensions (the
ones that seemed to be more easily recognisable in a short-term interaction): (1) Individualism
vs Collectivism and (2) Power Distance. As such, the other dimensions are left as future work.
3.3.1 Goal Utility
Goal utility is defined as a function that receives a goal and returns a numeric value that
indicates how much that goal is useful for the character (based on his current beliefs). Since
3.3. CULTURAL DIMENSIONS 29
beliefs are always changing on a dynamic environment, it is very likely that a utility of a
certain goal increases or decreases over time. For example, the goal of eating food has a high
utility when the character believes that he is hungry, even higher utility if he believes he is
starving or almost zero utility after he eats a satisfying amount of food. Rational characters
will continuously calculate the utility of every achievable goal and then select to focus on
achieving the goal with the highest utility at the moment.
Impact of Individualism Dimension
So, how can culture affect goal utility? Hofstede states that, in an individualistic culture, ”peo-
ple are expected to be only responsible for themselves and their immediate family.” [24] Also,
close friendships are very important. On the other hand, in a collectivistic culture ”everyone
looks out for one another in exchange for unquestioning loyalty”. As such, it seems clear that
our cultural characters should evaluate a goal’s utility under two different perspectives: (1) the
impact the goal has to themselves and (2) the impact the goal has to others (which requires
the ability to form mental models of others, like the agents from PsychSim). Individualistic
characters are much more concerned with the first perspective as the second one is only im-
portant if the character has a strong interpersonal attraction (symbolising a close bond) with
any of the other characters. Oppositely, collectivistic characters are equally concerned with
both perspectives and treat everyone alike (regardless of social bonds). Hence, we propose the
following equation (3.1) for calculating a goal’s utility based on the individualism score (IDV),
the impact the goal has on the character’s self (SI), the impact the goal has on others (OI),
and a positive relationship factor (PREL), which considers interpersonal attractions between
the targets of the goal and the character:
Utility(g) = SI(g) + OI(g)(100− IDV
100+
IDV
100× PREL(g)) (3.1)
Note that PREL(g) is normalised to a scale of 0 (no positive relationships) to 1 (maximum
positive relationships) and the exact equations for SI(g) and OI(g) are domain-dependent. To
explain the rationale behind this particular equation, the situation depicted in Figure 3.1 will
be used.
In this example, character A is considering the goal of giving some food to character B
versus the goal of giving some food to character C. A has plenty of food so loosing just a
little has a small negative impact, such as SI(g) = -1. However, B is hungry and poor, so
30 CHAPTER 3. CONCEPTUAL MODEL
receiving some food would have a considerable positive impact like OI(g) = 5. On the other
hand, C is also hungry but wealthy, so the impact for him of receiving some food is a little
lower, for example OI(g) = 4. Moreover, A has a negative interpersonal attraction towards B,
thus PREL(g) = 0. On the other hand, A has a positive interpersonal attraction towards C,
which makes PREL(g) return a positive multiplier depending on the intensity of the relation
(in this particular scenario, we’ll assume that it returns 0.5).
Figure 3.1: Giving Food Example
Using the previous situation, let’s examine three different cultural scenarios: (1) an extreme
collectivistic culture; (2) an extreme individualistic culture and (3) a neutral culture. In the
first scenario IDV is equal to zero and so equation 3.1 is reduced to:
Utility(g) = SI(g) + OI(g)
In this scenario, both goal impact functions are weighted equally which means that a
character considers his own well-being to have the same importance as the well-being of others,
regardless of the existent relationships. As such, regarding the example depicted, the utility
of giving B food is higher (Utility(g) = 4 ) than giving it to C (Utility(g) = 3 ).
For an extreme individualistic culture (IDV is equal to 100), the equation 3.1 changes to:
Utility(g) = SI(g) + OI(g)× PREL(g)
3.3. CULTURAL DIMENSIONS 31
In this scenario, the others well-being depends only on the existence of a positive relation-
ship. Since in the previous situation A disliked B, then PREL(g) = 0. Thus, A now will never
create an intention to give B food, since the goal has a utility of -1. But for C, since A has a
positive relation with him it makes PREL(g) return a positive multiplier (e.g. 0.5). Thus, the
utility of giving C food will now be equal to 2.
In the third scenario, a neutral culture (one that is neither inclined to individualism or
collectivism), the equation changes to:
Utility(g) = SI(g) + OI(g)× 0.5 + OI(g)× PREL(g)× 0.5
In this culture, the utility for giving B food is equal to 1.5. It is not negative but is lesser
than the utility of giving it to C, which is equal to 2. This means that generally characters of
a neutral culture care for all other agents but will give preference to their friends.
Impact of Power Distance Dimension
According to Hofstede [24], in low-power distance cultures people tend to regard others as
equals despite their formal status. Oppositely, in high power distance cultures powerful people
are expected to be privileged. As such, we want characters that belong to a high power culture
to favour goals that positively affect others who have a higher status. To achieve this result, we
propose to augment equation 3.1 with an additional fraction at the end, based on the power
distance score (PDI), and a power distance factor (DIST) that considers the differences of
power between the targets of the goal and the character:
Utility(g) = SI(g) + OI(g)(100− IDV
100+
IDV
100× PREL(g) +
PDI
100×DIST (g)) (3.2)
Similar to the positive relationship factor (PREL), DIST is also normalised to a scale of 0
(power equal or lower than self) to 1 (power is higher than self). Let’s consider again the goal
of character A (power = 5) to give food to character B (power = 3) and the goal of character
A to give food to character C (power = 10). Before, when we looked what would happen in
the extreme collectivistic scenario (IDV = 0), we concluded that A would prefer to give the
food to B (Utility(g) = 4) than to give it to C (Utility(g) = 3).
Now we’ll analyse how the situation changes with the addition of the Power Distance
dimension. Like before, we will examine three different scenarios: (1) extreme low power
32 CHAPTER 3. CONCEPTUAL MODEL
distance culture; (2) extreme high-power distance culture and (3) neutral culture. In the
first scenario PDI is equal to zero, which makes the DIST factor irrelevant, thus the previous
situation remains unchanged. In the second scenario, PDI is equal to 100, which reduces
equation (4.3.5) to:
Utility(g) = SI(g) + OI(g) + OI(g)×DIST (g) (3.3)
Since B has lower power than A, DIST(g) is equal to zero and so the goal of giving him
food remains with an utility of 4. Now, C has a power that is two times higher than the power
of A, thus DIST(g) will return a value greater than zero (e.g. DIST(g) = 0.5). Thus, the goal
of giving food to C has now an utility of 5 and so A prefers to give him the food instead of
giving it to B. Finally, in the neutral scenario (PDI = 50), the DIST(g) factor is divided in
half. This reduces the utility of giving C food to 4. Thus, A will not have any preference in
choosing to which agent it should give food, since both goals have the same utility in a neutral
power-distance and extreme collectivistic culture.
3.3.2 Emotional Appraisal
The idea that emotions are elicited by evaluations (appraisals) of events or situations, was first
introduced by Magda Arnold in [2]. Since then, different appraisal theories that attempt to de-
scribe the structure and/or the process of appraisal have been proposed (a detailed description
of several theories can be found in [52]). Even though the theories discord on several aspects,
they are all based on the idea that emotions result from a subjective evaluation of events. This
explains why two different people can show dramatic differences in their emotional response
to the same event.
But how does culture affects emotions? In the literature review, by Batja Mesquita and
Nico Frijda [42], ”cross-cultural differences as well as similarities have been identified in each
phase of the emotional process.” Regarding cultural differences that we can relate to the ap-
praisal process and to Hofstede’s dimensions, there are distinctions related to the Individualism
dimension, proposed by Markus and Kitayama in [40]. They argue that in individualistic cul-
tures the individual ”appears as focused on his or her independence and self-actualization”,
while in a collectivistic culture the individual is ”focused predominantly on his or her rela-
tionship with in-group members or with the in-group as a whole.” Consequently, individualists
appraise events in ”terms of their individual achievements and properties” while collectivists
3.3. CULTURAL DIMENSIONS 33
appraise events in ”terms of group the person belongs to or as affecting the interpersonal re-
lationships.” Note we did not find any distinctions that we could correlate directly to Power
Distance or any other dimension.
Based on the notions previously presented, we propose equation (3.4) for calculating the
praiseworthiness of an event. As stated in the OCC theory of emotions [46] (one of the most
used theories for synthesising emotions in agents), events with a positive praiseworthiness
will potentially cause the character to feel pride if he was responsible for the event, or feel
admiration towards the character that was responsible. On the other hand, a negative praise-
worthiness result will potentially cause the character to feel negative emotions such as shame
or reproach (depending also on who was responsible).
Praiseworthiness(e) =
0, if AI(e) > OI(e) ≥ 0
(OI(e)−AI(e))× 100−IDV100 , if otherwise
(3.4)
The equation we propose is based on the impact the event has on the character who caused
it (AI), the sum of impacts the event has on the other characters (OI), and the individualism
score (IDV). In general terms, the first branch of the equation refers to events that didn’t harm
others (OI(e) ≥ 0) but had a more beneficial effect for the character who caused them (AI(e)
>OI(e)). As such, no matter how collectivistic a culture is, a character will not be ashamed
if, for example, he has just eaten an apple (an event that had a positive effect on himself
but a neutral effect on others). As for the second branch, it provides the following results:
(1) the more collectivistic a culture is (i.e. the lower the IDV), the more an event that is
undesirable for others (OI(e) <0) but is beneficial for the responsible character (AI(e) >0) will
be blameworthy (e.g. stealing something); and also (2) the more collectivistic a culture is, the
more an event that is good for others (OI(e) >0) but is bad for the responsible character (AI(e)
<0) will be highly praiseworthy (e.g. giving food). In other words, collectivistic characters
will find highly admirable a spirit of self-sacrifice for the well-being of the group and will find
highly reproachable selfish acts.
To give an example, consider the following situation: agent B has asked directly agent A
for some food and A denies it. This has a positive impact on A considering he keeps the food
for future use (e.g. AI(e) = 1). However, it has a negative impact on B who is very hungry
(e.g. OI(e) = -3). Let’s consider that agent A and agent B are from a culture that has an IDV
34 CHAPTER 3. CONCEPTUAL MODEL
of 27 (the value found for the Portuguese culture). Applying the equation, agent’s A decision
will have a praiseworthiness value of -3 approximately. This means that A will potential feel
ashamed, while B would feel reproach for A. Instead, if A decides to give B food, it will have
a negative impact on A (e.g. AI(e) = -2) but a positive effect on B (e.g. OI(e) = 3). The
praiseworthiness value of this decision will be 3.6. As such, A will likely feel pride, while B will
feel admiration for A. Finally, if we re-examine both decisions, now considering the characters
belong to a culture with an IDV of 91 (the value of the USA culture), we’ll confirm that both
decisions have a very low praiseworthiness. Namely, the decision of giving B food will be equal
to 0.4, while keeping it -0.3.
3.4 Symbols
Hofstede defines symbols as ”gestures, words, pictures or objects that are given a special
meaning by the culture.” [24]. We decided to focus on gestures, as we believe that they are
the most important for our model, specially since they are used frequently in rituals.
As such, symbols are defined as a 2-tuple 〈N, M〉 where:
• N is the name of an physical action.
• M specifies the meaning of the physical action in that culture.
So, the main characteristic for a symbol is that it distinguishes a physical action (e.g.
waving hand) from its meaning (e.g. saying goodbye). Therefore, depending on their culture’s
set of symbols, characters can have different interpretations of the same action or perform the
same intention in different ways. Note that, to avoid ambiguity, the mapping between physical
actions and meanings is a one-to-one relationship, i.e. a particular action has only one meaning
and vice versa. We assume this is a simplification of the real world, where in fact the same
physical action can have different meanings in the same culture due to different contexts (e.g.
a bow can be a form of greeting but also a sign to acknowledge the applause for performing a
play).
Our motivation for focusing on gestures comes also from the fact that, as clearly stated
in [1], gestures are ”definitely NOT a universal language”, and it is frequent that unaware
outsiders can inadvertently offend someone by using the culturally ”wrong” gesture. A famous
example is when former American President George Bush greeted a large crowd of Australians
3.5. RITUALS 35
with the palm of his hand facing him, and his index and middle finger stretched out forming
a ”V”. He assumed he was performing a ”victory” gesture, yet in fact he was insulting the
Australians with a ”screw you” gesture.
As such, we believe gestures to be very important to an IVE that aims to be a successful
intercultural training tool. In fact, in our related work review, we discovered that gestures were
considered as one of the main cultural elements in the Tactical Language Training System [29].
3.5 Rituals
Hofstede defines rituals as essential social activities that are carried out in a predetermined
fashion. Yet, this is not a consensual definition. Since the earliest tribal communities hu-
mans have been involved in ritual activities. According to [7] rituals not only regulate the
relationships between one another in a community but also between people and their natu-
ral resources. In general a ritual can be defined as a set of actions, often thought to have
symbolic value, and its performance is usually prescribed by a religion or by the traditions
of a community. Although this definition seems straightforward, it does not define what kind
of activities make up for a ritual and in some extreme cases [15, 35], every activity can be
seen as a ritual. However, we do not consider this to be an interesting approach to rituals.
According to multiple authors[20, 36, 6, 38, 58], activities can be separated into two classes:
ritual activities and technical activities. Whilst a ritual activity is described as expressive,
rule-governed, routinised, symbolic, or non-instrumental, a technical activity is described as
pragmatic, spontaneous, and instrumentally effective. Therefore, we focus on rituals as a set
of ritual activities.
Another characteristic of rituals (one of the most important) is their invariance [7], in the
sense that a ritual is a repetitive and disciplined set of actions marked as precise and without
much invariance in them. In this perspective, a ritual can be understood as a recipe of activities
that should be executed in a predetermined way. This notion strongly resembles plan recipes
used in traditional BDI architectures [23], the difference being that traditional plans are based
on technical activities (the focus is the end result), whilst rituals are based on ritual activities
(the focus is in the sequence of steps). Thus, looking at how activities are represented in
traditional planning can help us structure a ritual. For instance, Rickel et. al. [51] consider an
activity to consist of ”a set of steps, each of which is either a primitive action or a composite
36 CHAPTER 3. CONCEPTUAL MODEL
action. Composite actions give tasks a hierarchical structure” and that ”there may be ordering
constraints among the steps”.
However, a ritual is more than a plan recipe because in addition to specifying how to be
executed, it must also specify when it should become activated and form an intention (similarly
to a goal). Hence, a ritual r is then formally defined as a tuple 〈T, R,C, S, O〉, where:
• T - specifies the type of the ritual. It associates each ritual with a name. Note that it
is necessary to define the type of the ritual because there might be several instantiations
of a given ritual type. For instance, a high-power culture may specify two different (or
even more) greeting rituals, one used between characters with low-status and the other
used by a low-status character to greet a high-status character. Although the actions
involved are different, the rituals performed have the same semantic (greeting someone).
• R - specifies the set of roles of participants involved in the ritual.
• C - represents the rituals’ context of activation, and it is composed by a set of conditions
that need to be verified in order for the ritual to be performed. These conditions must
also indicate the characters that may fit each of the specified roles.
• S - corresponds to the set of steps of the ritual, where a step is a pair 〈role, action〉. A
ritual usually involves the actions of other characters, thus it is necessary to define who
should perform each of the rituals’ actions.
• O - is the set of ordering constraints (if any) between the steps of the ritual. An order
constraint S1 ≺ S2 specifies that step S1 should be executed before step S2 starts.
To see how this definition can be applied, let’s consider an example on how to define a very
informal ritual of greeting someone:
• T - Greeting-Ritual.
• R - {Greet-Initiator, Greet-Replier}
• C - {c1,c2}, where c1 verifies if the character with the first role has just seen the character
that has the second role and c2 verifies if both characters have the same status.
• S - {<Greet-Initiator,Informal-Greet-Gesture>, <Greet-Replier,Informal-Greet-Gesture>}
3.6. CONCLUDING REMARKS 37
• O - {} There is no need for order constraints, since it does not matter which character
performs the first action.
To exemplify the addition of a formal greeting ritual to the same culture, we can define
another ritual with the same type as the previous one:
• T - Greeting-Ritual.
• R - {High-Power, Low-Power}
• C - {c1,c2}, where c1 verifies if the character with the first role has just seen the character
that has the second role and c2 verifies if the High-Power character has a higher status
than the Low-Power character.
• S - {<High-Power,Formal-Greet-Gesture>, <Low-Power,Respectful-Greet-Gesture-Reply>}
• O - {<High-Power,Formal-Greet-Gesture>≺< Low-Power,Respectful-Greet-Gesture-Reply>}
This constraint obliges the Low-Power character to wait first for the greeting gesture of
the High-Power character.
3.6 Concluding Remarks
In this chapter, we have presented a model to specify and create different cultures of au-
tonomous synthetic characters. Inspired by anthropological studies on human cultural vari-
ation, we proposed a definition of culture that encompasses three important elements: (1)
Cultural Dimensions, (2) Symbols and (3) Rituals.
Cultural Dimensions influence the behaviour of the characters by affecting their emotions
and the utility of their goals. Symbols are represented by physical actions that have a special
meaning only to the members of that culture. Finally, a ritual was defined as a well determined
sequence of actions, performed by specific roles when a specific context is verified.
38 CHAPTER 3. CONCEPTUAL MODEL
Chapter 4
Implementation
4.1 Introduction
This chapter describes an implementation of the conceptual model previously presented. The
implementation was done by extending an already existent autonomous agent architecture,
FAtiMA-PSI [17, 18, 37]. As such, we will start by presenting an summary of this particular
architecture also explaining why it was a good starting point for our work. We then present the
architecture that resulted from integrating our conceptual model. Finally, the implementation
of each cultural component is carefully described.
4.2 FAtiMA-PSI (Baseline Architecture)
FAtiMA-PSI (see Figure 4.1) is an agent architecture, written with the Java programming lan-
guage, which aims to create synthetic characters that are autonomous, engaging and believable.
In order to achieve this, emotions, needs, and personality take a central role in influencing be-
haviour. The concept of emotions is based on the OCC cognitive theory of emotions [12],
which defines emotions as valenced (good or bad) reactions to events. The individual evalua-
tion of events that causes such reactions is called the appraisal process. Characters’ needs are
grounded on a psychological model of human action regulation called PSI [19]. This theory
proposes five different human needs, that motivate human behaviour:
• Energy - the need to consume resources (water and food) in order to live.
• Integrity - the need to maintain one’s well being, avoiding pain.
39
40 CHAPTER 4. IMPLEMENTATION
• Affiliation - the need to be accepted by others and become part of social groups. It might
explain why humans perform rituals and abide by social values.
• Certainty - represents the need of being able to make good predictions about the envi-
ronment as well as the consequence our actions have in it.
• Competence - is the need of being able to solve problems and finish tasks with success.
As for personality, the architecture does not follow any specific theoretical model. Instead,
it allows to manually define different roles that are then associated to characters. The idea is
that these roles are perceived by the users as different personalities. Their authoring involves
defining the following characteristics: (1) Emotional Thresholds - how hard it is for a character
to feel a given emotion; (2) Emotional Decay Rates - how long does a character feels an emotion;
(3) Goals - the personal goals of the character; (4) Emotional Reaction Rules - how does the
character evaluates events; (5) Action Tendencies - reactive actions that are triggered by a
certain emotion (e.g. crying when feeling distress); and (6) Need Weights - the importance
each need has for the character.
Figure 4.1: FAtiMA-PSI Agent Architecture
4.2. FATIMA-PSI (BASELINE ARCHITECTURE) 41
The architecture, in general terms, works in the following way. Events that happen in the
environment are perceived by the agents’ sensors. Whenever a new event is perceived, the
following happens:
• The motivational state is changed according to the action’s impact on the character’s
needs (e.g. the eat action lowers the energy need). When an event lowers/raises the
agent’s needs, it is evaluated as desirable/undesirable for that agent. Moreover, the more
a certain need is low/high, the more higher/lower the utility of a goal that contributes
positively for that need is.
• The Autobiographic Memory and the Knowledge Base are updated. These are the ar-
chitecture’s main memory components. The first stores events and associated emotional
states while the second is responsible for storing semantic knowledge such as properties
about the world and relations between the characters.
• The reactive layer and deliberative layer appraises the event and generate a set of emo-
tions that alter’s the character’s emotional state. The deliberative appraisal elicits emo-
tions that are associated to the relation between the event and the agent’s goals. For
example, after a character completes a goal, he likely feels Satisfaction. On the other
hand, the reactive appraisal elicits all other types of emotions, based on how the event
affected the motivational state and also on manually defined emotional reaction rules,
which specify the desirability for other characters and the praiseworthiness of an event.
In both appraisals, the mood of the character and his emotional thresholds are also
considered. The resultant emotional state can possibly trigger action tendencies (quick
emotional reactions). These reactions are performed in the virtual world through the
agents’ effectors.
• New events trigger the goal selection process. This process starts by checking which
goals have become active (the ones that have all their preconditions satisfied). Then
for each active goal, if its expected utility surpasses a predefined minimum value, a new
intention to achieve the goal is created and added to the intention structure. Afterwards,
a continuous partial order planner, tries to build a plan for the intention with the highest
utility (the one associated to the goal that is expected to contribute the most for the
character’s needs). After a feasible plan is built, the planner executes it by sending the
42 CHAPTER 4. IMPLEMENTATION
actions to the effectors. Since the virtual world is a dynamic environment, the planner
is always monitoring the events that happen in the virtual world, to check if the current
plan needs to be changed or even dropped.
FAtiMA-PSI was chosen for the implementation of our cultural model, primarily because
it was specifically designed for the creation of agents that are autonomous synthetic characters
and resemble humans in the ways that they behave and interact with each other (agents have
emotions, quick reactions to events, goals, social relations, memory and needs). Moreover, the
architecture has the following processes that are necessary to implement our model: (1) goal
selection based on a goal utility function; (2) emotional appraisal; and (3) planning capabilities.
Finally, designing a new agent architecture from scratch would be too much time-consuming
and it is not the focus of our work.
4.3 Cultural Architecture
The integration of the cultural conceptual model into FAtiMA-PSI resulted in the architecture
depicted in Figure 4.2:
Figure 4.2: Cultural Agent Architecture
4.3. CULTURAL ARCHITECTURE 43
The new modules that resulted from this integration are: (1) Symbol Translator ; (2);
Motivational State of Others; (3) Cultural Appraisal ; (4) Cultural Goal Selection; and (5)
Ritual Manager. There is also a cultural parametrisation of Symbols, Dimensions and Rituals.
This parametrisation is done in a XML file that is associated to all the characters we wish to
belong to a specific culture. In the following sections we will discuss how each of the added
modules work as well as the details of the cultural parametrisation.
4.3.1 Symbols Parametrisation
Every symbol requires the manual definition of the parameters shown in Table 4.1.
Parameter Description
Name A string with the name of a physical action
Meaning A string with the meaning of the physical action
Table 4.1: Symbol parameters
The following XML code exemplifies the definition of two symbols:
<Symbol name="bow" meaning="respect-greeting-gesture"/>
<Symbol name="wave-hand" meaning="casual-greeting-gesture"/>
4.3.2 Symbol Translator
The Symbol Translator handles the specified symbols. Whenever an event is perceived by
the sensors, the module checks if the event’s action is associated with some symbol. If an
association is found, it simply changes the event by replacing the action name with its cultural
meaning. If no association is found, the event remains unchanged. To give an example, consider
the following event: agent A performs a bow to agent B. If in the culture file associated to
agent B the following symbol exists:
<Symbol name="bow" meaning="respect-greeting-gesture"/>
Then agent B will re-interpret the event as: agent A performs a respectful greeting gesture
to agent B. If no such symbol is found, the event is not re-interpreted. Also, whenever the
character wants to perform an intention that corresponds to a symbol’s meaning, the Symbol
44 CHAPTER 4. IMPLEMENTATION
Translator translates the intention to the associated physical action and sends it to the agent’s
effectors. For instance, if agent B intends to reply by performing also a respectful greeting
gesture to agent A, the symbol translator will translate this intention into it’s corresponding
action, causing the agent’s effectors to send a ”bow” action.
To optimise the translation process, while sacrificing some storage space, symbols are stored
simultaneously in two different hash tables: one that is indexed by the name of the physical
action and another that is indexed by the name of the symbol’s meaning. This was possible
due to the fact that we assume a one-to-one relationship between actions and meanings.
4.3.3 Motivational State of Others
When an agent is created, his own Motivational State is initialised. The initial intensity and
their relative weights are predefined in the character’s role file. This intensity can range from
0 to 10 (with the value 10 meaning that the need is entirely fulfilled and 0 meaning the need
is entirely unfulfilled). When an event happens, the associated action can have a predefined
effect that is used to update the motivational state. For example, if agent A eats an apple (an
action with a defined effect of 2 in the energy drive), he fulfills his energy level by 2 multiplied
by the weight (ranges from 0 to 1) the agent has for the energy need. Note that needs also
decay naturally with time.
In the new cultural architecture, agents are capable of reasoning not only about their
own needs, but also of other agents’ needs as well. To achieve this, every time an agent
encounters another agent he creates a mental representation of that agent’s motivational sate.
This representation is initiated with a neutral intensity for each drive (i.e. intensity = 5). This
default assumption is made since there is no evidence to decide if the other character has his
needs fulfilled or not. When other agents participate on events, the representation of their
motivational states is updated according to the event’s predefined effects on the affiliation,
integrity and energy needs (the other needs are only related to their own personal goals, which
are unknown to the agent).
Figure 4.3 was taken after running a small simulation where agent B meets and greets agent
A (an action that has a positive effect on A’s affiliation). The figure displays the resultant
data structures for their motivational states at the end. We can see that their perspectives
of each others’ motivational states are different from the real values. For example, agent B
thinks that agent A has his affiliation need at 7.7 while, in reality, agent A has only 5.7. This
4.3. CULTURAL ARCHITECTURE 45
Figure 4.3: Motivational States of Two Different Agents
happened because, before the greeting, agent B assumed A’s affiliation was at 5, not knowing
that in fact, A’s affiliation was much lower.
As we will describe later in this chapter, these mental representations of other agent’s
motivational states are fundamental for the Cultural Appraisal and Cultural Goal Selection
modules.
4.3.4 Dimensions Parametrisation
The parametrisation of the cultural dimensions is very simple, as described in Table 4.2. To
give an example, an extreme individualistic and high power distance culture is defined in the
culture XML file with the following code:
<CulturalDimension name="PowerDistance" value="100"/>
<CulturalDimension name="Individualism" value="100"/>
46 CHAPTER 4. IMPLEMENTATION
Parameter Range Description
Name{Individualism, Pow-
erDistance}The name of the dimension.
Value [0..100] The value associated to the dimension
Table 4.2: Cultural dimensions parameters
4.3.5 Cultural Goal Selection
To better explain how the Culture Goal Selection module works, we will start by describing
how goals are defined and treated in the architecture.
Goal Definition
All goals are predefined in an XML file, conventionally named Goal-Library. A goal can then be
associated to a single character, by including a reference in that character’s role file. However,
it can also be associated to every character of a given culture, by including a reference in that
culture’s XML file. To define a goal, one must author the following attributes:
• Id - the goal identifier or name.
• Preconditions - a list of conditions that determine when the goal becomes active.
• SuccessConditions - a list of conditions used to determine if the goal is successful.
• FailureConditions - a list of conditions that determine the goal failure.
• ExpectedEffects - specifies the expected effects the goal will have in the agents’ needs.
The Id is used to uniquely identify each instance of the goal, and it is useful to detect if a
given goal is already active or to search in the Autobiographic Memory for past activations. The
preconditions are used for the activation process; a goal becomes active if all the preconditions
of the goal are verified.SuccessConditions represent the state of the world that the goal aims to
achieve. Failure conditions represent an automatic goal failure mechanism. If at any time the
failure conditions are verified, the goal is assumed to fail and is removed from the deliberative
4.3. CULTURAL ARCHITECTURE 47
layer. The attribute ExpectedEffects, as we will see, is used in determining the goal’s expected
utility.
The following XML code is an example of a simple goal definition:
<ActivePursuitGoal name="OfferHelpToPaintNewHouse([target])">
<PreConditions>
<RecentEvent occurred="True" subject="[target]" action="SpeechAct" target="[SELF]"
parameters="ask-for-help-to-paint-new-house" />
</PreConditions>
<SucessConditions>
<RecentEvent occurred="True" subject="[SELF]" action="SpeechAct" target="[target]"
parameters="offer-help-to-paint-new-house" />
</SucessConditions>
<ExpectedEffects>
<OnSelect drive="Energy" target="[SELF]" value="-1"/>
<OnSelect drive="Affiliation" target="[target]" value="+3"/>
</ExpectedEffects>
</ActivePursuitGoal>
This goal is activated when another agent asks for help to paint his new house. The goal
is achieved just by saying that help will be given. The predefined expected effects of this goal
are a decrease on the agent’s energy level and an increase on the affiliation level of the agent
who asked for help. It’s important to note that this particular goal in the baseline architecture
would never be selected since there is no benefit for the agent who achieves it. However, with
the new cultural architecture, we will see that it can now become selected, depending on the
cultural dimensions parametrisation.
Goal Selection
The Cultural Goal Selection module starts by checking which goals associated to the character
have become active. Then, for each active goal it calculates their expected utility. The goal
with the highest utility at the moment is selected. If the utility of the goal selected is high
enough, the agent commits himself to the intention of achieving that goal. To calculate the
utility of each goal, the module uses the following equation:
Utility(g) = SI(g) + OI(g)(100− IDV
100+
IDV
100× PREL(g) +
PDI
100×DIST (g))
48 CHAPTER 4. IMPLEMENTATION
The rationale of this equation was already discussed in the Conceptual Model chapter. To
implement it, the following associations were made:
• SI(g) corresponds to the sum of the expected effects the goal g has on the character’s
current motivational state. Each expected effected is multiplied by an urgency factor.
This factor is inversely proportional to the need’s current intensity level. In the previous
goal example, assuming the urgency factor is equal to 1, then SI(g) = -1.
• OI(g) is the sum of the expected effects the goal g has on the goal’s target (determined
using the representation of that agent’s motivational state). It also uses the same urgency
factor. In the previous goal example, assuming the urgency factor of the target agent is
equal to 1, then OI(g) = 3.
• IDV corresponds directly to the individualism dimensional score, specified in the culture’s
XML file.
• PDI corresponds directly to the power distance index value, specified in the culture’s
XML file.
• PREL(g) is determined by searching in the Knowledge Base for a positive Like relation-
ship between the character and the target of the goal g; its value is normalised to a scale
of 0 (neutral relationship) to 1 (strongest relationship). Note that the Like relationships
are predefined in each character’s role file.
• DIST(g) is determined by subtracting the [self ](power) property from the [target](power)
property (after searching them in the Knowledge Base). If the result is positive, then
DIST(g) is normalised to a scale of 0 (no difference) to 1 (highest difference). If the
result is negative (means the target has a lower power), DIST(g) is equal to 0. Note that
the power properties for each character are also manually specified.
Using this equation, the previous goal of offering help to paint a house will likely have a
high utility if: (1) the agents are in a highly collectivistic culture (low IDV), or (2) if they are
in an individualistic culture (high IDV) and the agent has a strong relationship with the agent
who asks for help (high PREL), or (3) if the agent who asked for help has a higher status (high
DIST) and they are from a high-power distance culture (high PDI).
4.3. CULTURAL ARCHITECTURE 49
4.3.6 Cultural Appraisal
The Cultural Appraisal module is triggered by the perception of an event. It starts by deter-
mining the following appraisal variables:
• Desirability - represents how much the agent found the event desirable for himself. It is
equal to the impact the event had on the agent’s motivational state.
• DesirabilityForOther - represents how much the agent found the event desirable for the
other agents. It is equal to the perceived impact the event had on the other agents’
motivational states.
• Like - represent how much the agent likes/dislikes another agent or object. It is only
used when the agent perceives a new agent or a new object.
• Praiseworthiness - represents how the agent evaluates the event according to his cultural
values. This is determined by using the same equation we presented in the Conceptual
Model chapter.
Praiseworthiness(e) =
0, if AI(e) > OI(e) ≥ 0
(OI(e)−AI(e))× 100−IDV100 , if otherwise
OI(e) is equal to the impact the event had on the motivational state of the other agents
that were affected by the event. AI(E) is equal to the impact the event had on the
motivational state of the agent responsible for the event, IDV is equal to the individualism
dimensional score, specified in the culture’s XML file.
The values of these variables are used to determine a set of potential emotions and their
base intensities. For example, if an agent caused an event that had a Praiseworthiness of 3
then a new potential Pride emotion is created with a base intensity of 3 as well. Afterwards,
the agent’s current mood and arousal are applied to the base intensity to determine the final
emotion’s intensity. If the final intensity surpasses the agent’s defined emotional threshold for
that emotion, then it is added to the agent’s emotional state.
Note that the Cultural Appraisal module is an extension of the Reactive Appraisal of the
baseline architecture (described in greater detail in [16]). The main difference in the new
architecture is that it is capable of automatically calculating the praiseworthiness of every
50 CHAPTER 4. IMPLEMENTATION
event. In the baseline architecture, the Praiseworthiness value of each event had to be manually
specified in the character’s role file, in the form of emotional reaction rules. Still, in the new
architecture, the author can continue to define these rules in order to capture very particular
events that are also blameworthy/praiseworthy in the agents’ culture. For example, to define
as blameworthy the event of singing while eating at a dinner table.
4.3.7 Rituals Parametrisation
In the Conceptual Model chapter, we have said that a ritual is more than a plan recipe since in
addition to specifying how to be executed, it must also specify when it should become activated
and form an intention (similar to a goal). As such, the parametrisation of rituals (see table
4.3) has also some of the attributes found in goals.
Parameter Description
Name String with the ritual’s type identifier.
PreconditionsA list of general conditions that need to be verified in order to activate
the ritual.
ContextA list of conditions associated to the location, time and social rela-
tions that are also verified for the ritual’s activation.
Roles
A list of generic roles that are involved in the ritual. The ritual’s
preconditions and context dictate which character may fit a given
role.
StepsThe list of actions that have to be executed to finish the ritual. Every
action has an associated role that determines who should perform it.
Ordering Con-
straints
A set of links that specify the order in which the actions must be
performed through a set of links. A link simply indicates that a
particular step must be executed before another particular step.
Expected Ef-
fects
A list of effects the ritual will possibly have in the agent’s needs in
case the ritual succeeds
Table 4.3: Ritual parameters
4.3. CULTURAL ARCHITECTURE 51
The following XML code exemplifies a possible definition of a greeting ritual:
<Ritual name="GreetingRitual">
<Roles>
<Role name="[init]"/> <Role name="[replier]"/>
</Roles>
<PreConditions>
<RecentEvent occurred="True" subject="[init]" action="look-at"
target="[replier]"/>
<Property name="[replier]" operator="!=" value="[init]"/>
</PreConditions>
<Context>
<Social name="power" target="[init]" operator="=" value="[replier]"/>
</Context>
<Steps>
<Step role="[init]" name="casual-greeting-gesture([replier])"/>
<Step role="[replier]" name="casual-greeting-gesture([init])"/>
</Steps>
<OrderingConstraints>
</OrderingConstraints>
<ExpectedEffects>
<OnSelect drive="Affiliation" target="[SELF]" value="+3"/>
<OnIgnore drive="Affiliation" target="[SELF]" value="-3"/>
</ExpectedEffects>
</Ritual>
The ritual previously defined involves two characters, one that fits the init role, and the
other the replier role. The first precondition specifies that the ritual can be only activated if the
character who has the init role has looked at the replier. The second precondition states that
the same character can not fit both roles simultaneously. Afterwards, the context condition
indicates that the characters must have the same power. The ritual has a single step associated
to each role that involves the same action, casual-greeting-gesture. Since there are no ordering
constraints, it does not matter the sequence by which the actions are executed. Finally, the
expected effect for executing this ritual is a positive contribution to the affiliation need, while
ignoring it (choosing another ritual/goal instead), is expected to have a negative contribution.
52 CHAPTER 4. IMPLEMENTATION
Note that since the previous ritual only applies to characters that have the same power,
the architecture allows to define also a greeting ritual for characters of different power. This
new ritual can be very similar to the first (sharing the same type), changing just the context
condition to:
<Social name="power" target="[replier]" operator="LesserThan" value="[init]"/>
Naturally, since it is a different ritual it makes sense to also change the steps and ordering
constraints. A possible modification could be:
<Steps>
<Step role="[init]" name="casual-greeting-gesture([replier])"/>
<Step role="[replier]" name="respect-greeting-gesture([init])"/>
</Steps>
<OrderingConstraints>
<Link before="0" after="1"/>
</OrderingConstraints>
Hence, according to the new ritual, when a character greets another that has a higher
power he will have to wait for that character to first perform a casual-greeting-gesture so he
can then reply with a respect-greeting-gesture.
Differences from Goals
Unlike regular goals, a ritual does not need to define the success conditions and failure con-
ditions since they are implicitly defined. A ritual succeeds, according to a character’s point
of view, if all the ritual’s actions are perceived as successfully executed and fails otherwise.
There are another two important characteristics of rituals when comparing them with goals.
The first one is that the knowledge of the ritual is shared amongst all members of a given
culture, meaning that all agents will know about and have the same rituals. Contrarily, goals
are usually individual and the agents do not know what are the other ones’ goals. Secondly,
rituals with the same participants are considered equivalent if the roles are equivalent. As an
example, consider the previously defined greeting ritual, between character A and character
B, when (Power(A) = Power(B). The ritual of A greeting B is equivalent to B greeting A
since there is no distinction in the roles they take in the ritual (i.e both A and B can have the
init and replier role. Remember that from their individual perspectives, both A and B have
4.3. CULTURAL ARCHITECTURE 53
the same greeting ritual, and it may happen that A activates the ritual of greeting B (with A
assuming the init role) at the same time B activates the ritual of greeting A (with B assuming
the init role). By testing if a equivalent ritual exists before activating it, we can prevent agents
A and B from greeting each other twice.
4.3.8 Ritual Manager
This module is responsible for the ritual’s activation. A ritual can be activated by two distinct
processes: Pro-Active Activation and Reactive Activation. As for the the ritual’s execution, it
is done using the architecture’s Planner.
Pro-Active Activation
This pro-active activation is the main process of ritual’s activation, which is similar to the
goal’s activation process. When an event is perceived, the deliberative layer will check all
Rituals’ context conditions and normal preconditions to determine if they can become active.
This process also specifies who will be the possible participants of the ritual by looking at the
context’s conditions. If the context’s conditions are not enough to specify all the roles of the
ritual, it will not be possible to instantiate the ritual and thus it will never be activated.
However, even if the conditions are not enough, the deliberative layer will be responsible
for searching in the agent’s Knowledge Base (or in other knowledge components), in order
to find a set of substitutions that will make the conditions valid. If such sets are found, the
activation process will create different instantiations for each of the possible substitutions.
As a very simple example, imagine a Greeting ritual with two roles, [initiator] and [replier],
and with two conditions [initiator](type) = character and [replier](type) = character. The
activation process will check the Knowledge Base searching for entities that have a property
type with the value character, and will then create distinct ritual instantiations for each of the
valid combinations. Supposing that the agent knows two entities that satisfy such property, A
and B, he will create four active instantiations: Greeting(A,A), Greeting(A,B), Greeting(B,B),
Greeting(B,A). Obviously, the ritual must specify more conditions such as the initiator being
different from the replier, the greeting only happening when the initiator sees the replier, etc.
54 CHAPTER 4. IMPLEMENTATION
Reactive Activation
The second type of activation is the reactive activation. Although the pro-active activation
process handles most of ritual activations, there are some situations where it cannot activate
a ritual when it should. This happens for instance when someone else started the ritual before
the agent could activate it’s own ritual (or started it with a different role allocation). To
illustrate the situation, we can consider the above situation where agent A decided to take the
initiative to greet B but B was faster and took the initiative. Instead of continuing with the
original ritual, it should detect that he just need to continue the ritual started by B.
In order to handle these situations, a reactive process of ritual activation was integrated
in the architecture. When the agent perceives an action performed by another agent, it will
compare the action against all the rituals he knows. If the action belongs to a ritual, he will
then check the ritual’s preconditions to see if the ritual is valid in that situation. Finally if the
ritual is valid, the agent will also verify if he has an active role in the ritual or not. If he does
not has any role, there is no need to activate the ritual since he will not do anything anyway.
If, on the contrary, the agent has a possible role on the ritual, he will activate the ritual and
the ritual will then be executed as a regular ritual. This is similar to what humans do when
they recognize a ritual they know and that they should follow.
Execution
Once a ritual becomes active, the architecture’s Planner will pick up the predefined plan and
adjust it to the current state of the world (and may even add missing actions if the ritual is
incomplete). This means that the planner will remove actions of the ritual that were already
executed and will check if there are any preconditions necessary for the execution of the steps
(and not verified at the moment). If there are, the planner will try to add actions that achieve
the necessary preconditions. For instance, imagine that in order to perform a ritual action of
bowing to someone, the agent would need to move near that character. The planner would
detect such precondition and add the corresponding move action to the plan.
After this initial processing step, the planner will then start to execute the ritual if the plan
is valid, or will fail if there is no possible way to execute the plan. When selecting an action
for execution, the planner will always give preference to actions that must be executed by the
agent (this is important to prevent some types of deadlock). If such actions are available, and
4.4. CONCLUDING REMARKS 55
there is no ordering constraint that prevents them from being selected, the agent will send one
of them for execution. The agent will then monitor the execution of the action, and will select
the next action once the last one finishes. If the only available actions correspond to actions of
other agents, the agent will wait a predefined amount of time for the other agent to act. If after
some time, the other agent performs as expected the agent will update the ritual accordingly
and move to the next action. On the contrary, if time goes by and nothing happens, the agent
will lower its expectations about the ritual, and may eventually give up.
4.4 Concluding Remarks
The purpose of this chapter was to describe how our cultural conceptual model was integrated
into an agent architecture called FAtiMA-PSI. We have started by describing the main features
of this particular architecture and explaining why it was a good starting point for our work.
Afterwards, the extended architecture for culture-specific characters was presented. It embeds
the characters with a predisposition to prefer certain goals and feel certain emotions according
to their culture. The main idea for achieving this was to make characters more or less con-
cerned with the needs of others, according to the values defined for their culture’s dimensions.
Moreover, this required characters to form their own mental models of others’ motivational
states. Additionally, we developed a symbol translator component that captures specificities
of cultural communication by using a simple parametrisation of symbols. Finally, the archi-
tecture allows characters to perform rituals that are associated to their culture. These rituals
were implemented in a similar manner to how goals are implemented, yet with significant dif-
ferences, such as having an already predefined plan associated. Also, the ritual’s activation
and execution processes were carefully described.
56 CHAPTER 4. IMPLEMENTATION
Chapter 5
Case Studies
5.1 Introduction
This chapter describes two case studies that were developed to evaluate the architecture’s
capability of creating different synthetic cultures. The first case study consists in a serious
game called ORIENT where users interact with a group of Sprytes, a fictional culture of
synthetic characters driven by our architecture. The second case study consists in a short
emergent story - a dinner party - with four different cultural scenarios. The story is acted
by a small group of characters, which interact with each other in a virtual world where the
user is an invisible observer. In this chapter, we will explain the common elements of the four
scenarios, as well as their differences.
5.2 First Case Study - ORIENT Game
The implemented cultural architecture was used for the development of a serious game called
ORIENT: Overcoming Refugee Integration with Empathic Novel Technology. The game is
an agent-based educational role-play, developed in the context of an EU-funded project called
eCIRCUS. The main purpose of ORIENT is to promote inter-cultural empathy for young
teenagers. With that in mind, the progressive challenges the players must confront during the
game were designed based on Bennet’s developmental model of intercultural sensitivity [8] (a
model analysed in the Related Work chapter).
57
58 CHAPTER 5. CASE STUDIES
Figure 5.1: Screenshot of ORIENT.
In ORIENT (see Figure 5.1), players (assuming the role of space travellers), must interact
with Sprytes, an unfamiliar fictional foreign culture whose planet is about to be destroyed by a
large meteor. The main objective for the players is to gain the trust of the Sprytes to then save
them from annihilation. To gain their trust, players have to become familiar with the Sprytes
strange customs and gestures. For instance, they must understand that Sprytes’s culture is
strongly hierarchical and everyone is highly compassionate and loyal to each other. In order
to create this culturally specific behaviour, we applied our architecture to define the Spryte’s
gestures and rituals. Moreover, their culture was parametrised as highly collectivistic and with
a high power distance score.
ORIENT also has an innovative approach in terms of user interaction. Instead of only
one user interacting with a gamepad or a keyboard and mouse, ORIENT allows three users
to interact simultaneously, each one controlling one of the following devices: a Dance Mat, a
mobile phone and a WiiMote. Each device has a different but essential function: (1) the Dance
Mat is used for navigation purposes; (2) the mobile phone is used for verbal communication
and object recognition; and finally (3) the WiiMote is used to perform important cultural
gestures that are used for instance, in the greeting rituals of the culture.
The rationale for allowing a group of users to interact simultaneously was to promote
social collaboration. A second objective was to encourage discussion between players about
the cultural differences found in the synthetic culture. Also, the intention of using novel
5.3. SECOND CASE STUDY - DINNER PARTY 59
interaction devices was to incite players’ curiosity to play the game and to provide a more
engaging experience.
After conducting two pilot studies, users did found the Sprytes to be a very different culture
from their own and most users were interested in the storyline. However, even tough ORIENT
seems to be a promising project, it currently has only a single culture. Hence, we cannot use
it to measure the power of our architecture in creating distinct cultures. For that reason, we
developed another case study, with a small group of characters configured with different cultural
parametrisations. This second case study was the one used for our evaluation. Therefore it
will be described in greater detail in the following sections.
5.3 Second Case Study - Dinner Party
The design process for the second case study involved the choice of a simple and common
real-life situation that had significant cultural variation. The idea was to use our architecture
to create different cultural groups of autonomous characters to perform that situation in the
virtual world. Also, asides from a different cultural parametrisation, the scenarios would have
exactly the same configuration.
Our final decision was to have five different characters acting a simple dinner party together.
We have chosen this situation because cultures throughout the world have very different eating
rituals [59]. Therefore, it seemed a good situation to portray cultural differences with our
system. Also, for the connection between the agent’s architecture and the 3D virtual world,
we decided to reuse the same platform that was used in the ORIENT game [4].
For simplicity reasons, the overall plot is very short: the characters arrive at the party
location; greet each other; socialise for a while; and then sit together at a dinner table and
start to eat. Note that, even though the story follows a plot, it is not explicitly scripted. What
happens is that the story emerges from the characters’ autonomous decisions, which are based
on their goals, emotions, motivational needs, and culture. As authors of the story, we only
defined: (1) properties of the characters/environment; (2) the set of possible goals/behaviours
the characters might decide upon; and (3) a list of simple Narrative Actions such as inserting
a character in the environment at a given time (these actions are performed by a special entity
of the world platform, called Story Facilitator [22]). Before we dwell on the differences of the
scenarios, we will first describe the characters in more detail.
60 CHAPTER 5. CASE STUDIES
5.3.1 Character Design
Figure 5.2: Characters at the dinner table
All the scenarios have the same five different characters. Figure 5.2 shows all of them having
dinner. Noticeably, they all have the same appearance and are dressed with the same peculiar
outfit (just with different colours). The main reason for designing the characters in this manner,
was so that users could not associate the characters’ appearance to any particular known
culture. In our opinion, that would very likely lead the users to create cultural expectations
that would not be related to our cultural model.
Despite the fact that the characters all look alike, they have some individual differences.
Namely, two of them have a low social status (the ones dressed in red); another two have a
medium status (the ones with the blue clothes) and the last one has a high status (the elder of
the group, dressed in violet). Moreover, there is one character that feels sick and another that
has some medicine with him. Also, the character that has the medicine has just built a new
house and needs someone to help him paint it. This character also likes to tell ”lame” jokes
to everyone. In addition, there is one character that has a small gift to offer to the host of the
party, yet the host character dislikes small gifts. Finally, they all start the interaction with a
neutral Like relationship between each other.
All of these individual differences correspond to different properties that are stored in the
5.3. SECOND CASE STUDY - DINNER PARTY 61
Knowledge Base of all the characters. They were used as preconditions to activate certain goals
or rituals. As we will describe in the next section, the different social statuses were mainly used
for the specification of rituals, while the other differences were made for originating situations
to explore the parametrisation of the culture’s dimensional scores. Also, we considered that
giving some individuality to the characters was important for creating richer scenarios.
5.3.2 Culture Design
The choice of the different cultural scenarios, was based on the notion of synthetic cultures that
was mentioned in the Related Work chapter. They represent extreme manifestations of the
cultural dimensional values, and so they stress the occurrence of the associated behavioural
tendencies. This makes the cultural differences more likely to be recognised, which is very
important for our objective.
Hence, we’ve defined the following four different cultures, based on the extremes of the two
dimensions implemented:
Culture 1 - Extremely High Power Distance with Extreme Individualism:
<CulturalDimension name="PowerDistance" value="100"/>
<CulturalDimension name="Individualism" value="100"/>
Culture 2 - Extremely High Power Distance with Extreme Collectivism:
<CulturalDimension name="PowerDistance" value="100"/>
<CulturalDimension name="Individualism" value="0"/>
Culture 3 - Extremely Low Power Distance with Extreme Individualism:
<CulturalDimension name="PowerDistance" value="0"/>
<CulturalDimension name="Individualism" value="100"/>
Culture 4 - Extremely Low Power Distance with Extreme Collectivism:
<CulturalDimension name="PowerDistance" value="0"/>
<CulturalDimension name="Individualism" value="0"/>
62 CHAPTER 5. CASE STUDIES
In addition to this parametrisation, the four different cultures share a set of common goals,
yet have different rituals and symbols associated. We will now discuss the specification of these
elements in more detail.
5.3.3 Goals
Like described earlier, the dimensional values will have a significant effect on the goal selection
process. As such, 16 different goals were defined for this particular dinner party scenario. Note
that since every new ”physical” action requires a new graphical animation, the majority of the
goals defined only involved SpeechActs.
It is important to note that these goals are associated to every culture and thus are known
by every character. As an example, the character that has a new house which still needs paint-
ing will activate the goal AskForHelpToPaintNewHouse. Then, the character who is asked
for help will choose between two goals: (1) OfferHelpToPaintNewHouse and (2) DenyHelp-
ToPaintNewHouse. If the characters have a neutral Like relationship, the first decision is
more likely to happen in the extreme collectivistic cultures, while the second is more likely to
be chosen in the extreme individualistic cultures. This happens because the first goal requires
the character to spend his own energy to benefit the owner of the house. Remember that,
according to Hofstede, in collectivistic cultures people tend to always look out for one another
while in the individualistic cultures people assume that they are only responsible for those they
share a very close bond [25]. Finally, if the character asked for help has a much lower status
than the owner of the house, he will then be more inclined to offer his help in an extreme high
power distance culture.
5.3.4 Rituals
Since the plot of our scenarios is relatively simple, we found it hard to create different rituals
for each of the four cultures defined. As such we decided to focus only on the Power Distance
and created two sets of specific rituals to reflect the opposite extremes of this dimension. We
have chosen Power Distance over Individualism, because we believe it has a greater impact
on rituals (people from a high power distance culture are very formal and ceremonious, while
people from a low power distance culture are the opposite [25]).
In total, agents will perform three ritual types: (1) Greeting Ritual (according to [7] greet-
ings are patterned routines that can be seen as mini rituals); (2) Welcoming Ritual ; and (3)
5.3. SECOND CASE STUDY - DINNER PARTY 63
Dinner Ritual. The ritual differences that exist between the High Power Distance and the Low
Power Distance cultures are:
• Greeting Ritual - it is activated when two characters look at each other. In the Low
Power Distance cultures, the ritual is the same for all different status: the two characters
execute mutually a casual greeting gesture and say a casual greeting sentence to one
another. However, in the High Power Distance culture, rituals differ according to the
social status of the participants. When characters have the same status, they perform the
same actions as in the low power scenario. Yet, when a character greets a higher status
character, he has to bow to that character and also say a respectful greeting sentence,
which the character then replies with a casual greeting sentence. Furthermore, a different
ritual is used when greeting the elder. This ritual consists only in the character bowing
to the elder, while the elder doesn’t reply in any way.
• Welcoming Ritual - this ritual is performed by the host character when he welcomes the
guests to the party. The ritual consists simply in the host saying welcome to all agents
present, and then each agent replies with a gratitude sentence. The difference between
the Low Power Distance and the High Power Distance cultures is that the gratitude
sentences are far more formal and polite in the High Power Distance scenarios, while in
the Low Power cultures, the gratitude sentences are far more informal. Furthermore, the
host in a High Power Distance culture has to wait for the elder to arrive before starting
the ritual.
• Dinner Ritual - this ritual is activated after all the guests have arrived to the party. It
consists in the host announcing to the characters that the dinner will start and everyone
should take their seats. Then the ritual proceeds with the characters seating at the table
and starting to eat. However, while in the Low Power Distance cultures everyone rushes
to the table immediately, not even waiting for the host to finish the announcement (see
Figure 5.3), in the High Power Distance cultures everyone has to wait first for the elder
to sit before they can sit(see Figure 5.4), and then they have to wait for the elder to
start eating before themselves can eat. Moreover, the elder in the High Power Distance
culture has the privilege to sit in the more fancy chair.
64 CHAPTER 5. CASE STUDIES
Figure 5.3: Dinner Ritual (Low Power Distance Culture) - Screenshot Sequence.
Figure 5.4: Dinner Ritual (High Power Distance Culture) - Screenshot Sequence.
5.3.5 Symbols
Alongside the ritual definition, we also associated different symbols for the extremes of the
Power Distance dimension. Hence, the High Power Distance cultures have these two symbols:
<Symbol name="bow" meaning="respect-greeting-gesture"/>
<Symbol name="wave-hand" meaning="casual-greeting-gesture"/>
And the Low Power Distance cultures have only this symbol:
5.4. CONCLUDING REMARKS 65
<Symbol name="thumb-up" meaning="casual-greeting-gesture"/>
Note that the symbols created were all used in the specification of the greeting rituals.
Our rationale for selecting these specific symbols was to emphasise the differences between the
opposite extremes of the Power Distance dimension, based on symbols found in real cultures.
5.4 Concluding Remarks
In this chapter, we presented two different case studies where our agent architecture was
applied. The first case study consists in a game where users must be socially accepted by
a group of autonomous synthetic characters that have a very distinct culture. To make this
group of characters act in a cultural distinct manner, our agent architecture was used. The
second case study consisted in a short emergent story - a dinner party - with a small group of
characters, parametrised with four different cultures (using different dimensions, rituals and
symbols). The design of this emergent story was described, focusing on what lead us to choose
this particular situation. The design of the characters was also discussed. All the characters
have the same peculiar appearance to minimise cultural expectations that might arise in users.
Also, we gave individual properties to the characters to enrich the scenarios. In the remainder
of the chapter, we described the design of the different cultural parametrisations.
66 CHAPTER 5. CASE STUDIES
Chapter 6
Evaluation
6.1 Introduction
This chapter describes two different experiments that we have conducted using the Dinner
Party cultural scenarios of the previous chapter. Both experiments consisted in users visualis-
ing two cultural groups of characters with a different cultural parametrisation. Hence, we begin
with a discussion of the main objective of the evaluation, while the remainder of the chapter
focus on the two different experiments, describing their different methodology and results.
6.2 Objective
The work described in this dissertation focused on the following problem:
How can we build different cultural groups of autonomous synthetic characters that
exhibit distinguishable differences in their patterns of behaviour, similar to those
found in real human cultures?
In order to address this problem, we created an agent architecture with an explicit model
of culture, which was grounded on anthropological studies of behavioural patterns in human
cultures. As such, the objective for our evaluation was to determine if the cultural elements we
have implemented are sufficient enough to lead to the user’s perception of different cultures.
The reason for conducting two experiments was to be able to compare the effect of only
modifying cultural rituals (and their associated symbols) versus the effect of solely modifying
the cultural dimensions parametrisation. For these experiments, we decided not to use the
67
68 CHAPTER 6. EVALUATION
scenario with a low power distance and collectivistic culture, because we did not found any
real culture that fits into that category.
6.3 First Experiment - Rituals
Our hypothesis with the first experiment is that participants will be able to perceive cultural
differences in the behaviour of two groups of characters, just by associating a different set of
rituals to each group.
6.3.1 Methodology
The design for the first experiment consisted in running our system with two different cultural
scenarios from the ”dinner party” case study: one with the High Power distance rituals and
the other with the Low Power distance rituals (both are extreme individualistic). A video was
obtained for each one of these cases.
Both videos were then used in an online questionnaire (Annex A), which starts by asking
participants to watch one of the videos and then answer two groups of questions about the
characters depicted in it. Afterwards, the participants were asked to watch the other video
and again answer the same groups of questions. After seeing both videos, the questionnaire
consists in two additional questions that tries to access if any differences between the videos
presented were perceived, and if so, if participants understood those differences as being caused
by the culture of the characters, or by their personalities, or by neither one of these factors.
Finally we asked users their gender, age, and nationality.
Since repeated measures were used, participants were randomly assigned to a visualisation
order. Roughly half of them saw the Low Power culture video first, and the High Power
culture second. Also, the criteria used for selecting participants was: (1) a good knowledge of
the English language, and (2) no prior knowledge of our system and/or its purpose.
Regarding the different question groups of the questionnaire, the first one consisted in six
different statements (e.g. ”The eldest male should be the head of the household”) to which the
participants had to decide if they were appropriate to the characters or not. These statements
represent cultural values associated either to a High Power/Low Power/Individualist/Collectivist
culture. Furthermore, the statements were based on the questions used by Hofstede in his ques-
tionnaire. In the second question group, participants had to choose a number between two
6.3. FIRST EXPERIMENT - RITUALS 69
opposite adjectives in a scale from -3 to 3, according to what they thought to fit best with
the characters (if the participant thought neither the left nor the right adjective matched the
characters shown, he/she was asked to choose the zero value). The adjectives chosen were:
Approachable/Distant; Pleasant/Unpleasant; Unfriendly/Friendly; Relaxed/Tense; Compas-
sionate/Indifferent; Serious/Cheerful; and Warm/Cold.
6.3.2 Results
We had a total of 41 participants aged between 18 and 40 years old of which 73% were male.
Regarding nationality, 39 participants were Portuguese, 1 Italian, and 1 South African.
Regarding the questions about the value statements, the results were inconclusive. A
possible explanation was that the questions themselves were confusing. In fact, after the
evaluation was done, many users reported that they felt they could not answer properly about
some of the statements due to lack of information for a ”correct” answer. One such statement
was ”It is right for workers to openly disagree with their superiors”. However, our intention
with this question group was for participants to respond just intuitively based on their first
impression made by the video.
Considering the adjectives, the Wilcoxon test was applied because the data was not fol-
lowing a normal distribution (some of the questions presented binomial distributions). The
results obtained are shown in Table 6.1. For two pairs of adjectives, Pleasant/Unpleasant
and Approachable/Distant, the result was very far from being statistically significant (p=0,898
and p=0,709), which signifies that the cultural differences expressed had no influence in these
particular adjectives. Also, the results for the adjective pairs Relaxed/Tense and Compassion-
ate/Indifferent are not statistically significant but are close to it (p=0,07 and p=0,08), with
both having a small effect size (r = 0,20 and r=0,19). This only suggests that the Low Power
Distance culture is perceived as slightly more relaxed and compassionate than the High Power
Distance culture.
More interestingly, the Serious/Cheerful pair yields significant results (p=0,001) and thus
we can affirm that there is a significant effect of the cultural behaviour in the user’s classification
of these adjectives, with a medium size (r=0,38). This indicates that users found the Low Power
culture cheerful and the High Power culture serious.
In the last two questions to assess directly if users perceived the videos as being different,
only 4 did not found any differences (from the 41 participants). This corresponds to near 10%
70 CHAPTER 6. EVALUATION
Table 6.1: First Experiment - Results for the user’s adjective classification
of the participants. We applied a Chi-square test just to make sure the result was not obtained
by chance. The Chi-square value obtained was 26,56 and it was highly significant (p=0,000).
From the resulting 37 participants (which answered they had perceived differences), we asked
them if they thought the differences were related to the character’s culture, the character’s
personality or neither. 67% answered culture, 30% personality and only 3% answered neither.
We performed a similar chi-square test, but removing the only participant that answered
neither of them. The Chi-square value obtained was 5,444 and was significant (p=0,02). These
are very good results since they imply that most users did in fact find differences in the videos,
and most of them considered those changes to be caused by the character’s cultures.
6.4 Second Experiment - Dimensions
The objective of the second experiment was to determine if users could recognise cultural
differences in the behaviours of two groups of characters, solely by changing their associated
value for the Individualism dimension. Namely, we wanted to check if users did in fact could
recognise one group as more individualistic and the other as more collectivistic.
6.4. SECOND EXPERIMENT - DIMENSIONS 71
6.4.1 Methodology
The design of the second experiment was very similar to the first one, involving the same type
of questionnaire, but now of one the videos shows an extreme collectivistic culture and the
other an extreme individualistic culture (both cultures have rituals associated to a High Power
distance). Moreover, the questionnaire suffered some revisions (see Annex B). Due to the
criticisms about the questions for the values statements, we’ve done the following changes: (1)
altered the phrasing of the first question group to ”Regarding the group of characters shown in
the previous video, what do you feel about how they think and act?”; (2) added a note stating
”the idea is to give an opinion based only on your first impression of the characters”; and (3)
changed the answer options from yes/no to a scale from -3 to 3 (similar to the one used in the
adjectives). Furthermore, we have added three more questions about values (see Table 6.2),
and an extra five pairs of adjectives (see Table 6.3).
6.4.2 Results
For this second experiment we had a total of 42 participants aged between 18 and 34 years old
of which 76% were male. Regarding nationality, 36 participants were Portuguese, 5 German,
and 1 British.
Concerning the values question group, unlike the first experiment, we now have many
significant results (again we used the Wilcoxon test). In Table 6.2 we can see that the results
for every value statement related to Individualism or Collectivism is statistically significant
(p <0,05). As such, we can affirm that users found the individualistic values to be more
appropriate for the individualistic culture and the collectivistic values to be more appropriate
for the collectivistic culture. The highest effect (r=0,38) was for ”They like to trust and
cooperate with other people” statement. In our perspective, these are very good results.
They suggest users can recognise appropriate differences related to cultural values in groups of
characters, by simply changing their parametrisation of our dimensions component accordingly.
Also note that none of the values related to the Power Distance dimension yielded signif-
icant results (p >0,05). This signifies that the cultural differences expressed (related only to
Individualism) had no significant influence for these particular values. However, one of these
statements (”Power and wealth are bad”) was close to being significant (p=0,058), with a weak
effect (r=0,021).
72 CHAPTER 6. EVALUATION
Table 6.2: Second experiment - Results for the statements classification.
For the adjective’s classification we used the Wilcoxon test once more. The results are
shown in Table 6.3. Except for the Equal/Biased and Warm/Cool every other pair of ad-
jectives yields significant results. Thus we can affirm that there is a significant effect of the
Individualism dimension score in the user’s classification of most adjectives. Amongst them are
the adjectives Individualistic/Collectivistic (which has the largest difference in averages) and
Independent/Sharing. This constitutes a very good result, since it demonstrates that user’s
interpretation of the characters’ behaviour matches the parametrisation of the dimensions
component.
6.4. SECOND EXPERIMENT - DIMENSIONS 73
Table 6.3: Second experiment - Results for the user’s adjective classification.
Users also found the characters in the collectivistic video to be more: Approachable; Polite;
Pleasant ; Friendly ; and Compassionate. All these adjectives have a positive connotation, but
aren’t directly related to Individualism or Collectivism. One possible explanation for this result
comes from the fact that the majority of the participants is from Portugal, which is a strong
collectivistic culture according to Hofstede’s findings (see Table 2.1). Assuming most of the
participants are in the ethnocentric stages of the DMIS model (see Figure 2.1), it is plausible
that they would rate the behaviour of another collectivistic culture more favourably. As such,
we think it would be interesting to repeat this experiment with participants from a strong
74 CHAPTER 6. EVALUATION
individualistic culture, such as the USA.
Finally, note that there was also a significant effect for adjectives that are more related to
the Power Distance Dimension. For example, users found the collectivistic culture to be more
Hierarchical than the individualistic culture, yet both cultures had a High Power Distance.
This is an interesting result that suggests that the behaviours caused by one dimension can
alter the user’s perception of behaviours more directly related to another dimension.
In the last two questions to assess directly if users perceived the videos as being different,
only 1 did not found any differences (from the 42 participants). This corresponds to only 3%
of the participants. We again applied a Chi-square test just to make sure the result was not
obtained by chance. The Chi-square value obtained was 38,095 and it was highly significant
(p=0,000). From the resulting 41 participants (which answered they had perceived differences),
63% associated the differences to personality, 30% to culture and only 7% answered neither.
We performed a similar chi-square test, but removing the participants that answered neither.
The Chi-square value obtained was 5,158 and was significant (p=0,023).
6.5 Concluding Remarks
The purpose of this chapter was to describe the evaluation done to our agent architecture, using
the dinner party scenarios defined in the previous chapter. This evaluation consisted in two
different experiments, one related to Rituals and the other related to Dimensions. The overall
objective was to determine if the representation of these elements in our cultural architecture
is strong enough to lead to the user’s perception of different cultures.
Both experiments had a similar methodology and gave significant results. Regarding the
question where we ask explicitly if the users found differences in the groups visualised, most
users answered positively in both experiments. However, when asked about if those differences
were due to personality or to culture, the experiments had significant differences (see Figure
6.1). Even though less users found the differences were attributed to culture in the dimensions
experiment, it was in this experiment where users found the two groups to be more distinct,
in terms of adjectives and values.
This result corroborates Hofstede’s argument that different behavioural tendencies, associ-
ated to his dimensions are harder to interpret as cultural by the average person than rituals.
Also, note that despite the fact we exaggerated these tendencies by using an extreme parametri-
6.5. CONCLUDING REMARKS 75
Figure 6.1: Results for: Do you think the differences are related to the culture or personality?
sation, users only saw two small groups of characters interacting with each other in a short
period of time. It would be interesting to see what would happen if the scenarios were longer
with many more characters and situations.
76 CHAPTER 6. EVALUATION
Chapter 7
Conclusion
In this dissertation, we argued that culture, a fundamental aspect of human societies, is an im-
portant notion to consider when developing social intelligent agents to represent autonomous
synthetic characters. However, culture is a very complex concept that has been greatly over-
looked in current systems. As such, we hope the work presented in this dissertation constitutes
an important step towards the answer of the following question:
How can we build different cultural groups of autonomous synthetic characters that
exhibit distinguishable differences in their patterns of behaviour, similar to those
found in real human cultures?
Our hypothesis for solving this problem was:
If the behaviour of autonomous synthetic characters is driven by an agent archi-
tecture with an explicit model of culture, that is inspired by some aspects of human
cultures, users will be able to recognise cultural differences in different groups of
characters that differ exclusively in their cultural specification.
In order to prove this hypothesis we looked into the literature of anthropology to investigate
important elements of human cultures. Also, we conducted a small survey on autonomous
agents that were designed with cultural and social behaviour in mind. Inspired by our findings,
we defined a conceptual model of culture, which involved a dimensional model of behavioural
tendencies, rituals, and symbols. We then integrated this model into an already existent
autonomous agent architecture.
77
78 CHAPTER 7. CONCLUSION
With the intention of evaluating the implemented architecture, we created a case study with
different cultural scenarios. They revolve around the same small group of characters having
a dinner party, differing only in the group’s cultural parametrisation. For the evaluation, we
have decided to conduct two separate experiments to compare the effect of only modifying
cultural rituals (and their associated symbols) versus the effect of solely modifying the cultural
dimensions parametrisation.
For the first experiment, participants watched two cultures that only differed in their rituals:
an individualistic high power distance culture and an individualistic low power distance culture.
67% of the participants interpreted the differences as being caused by the characters’ culture
(a statistically significant result). This is a very encouraging result as it shows that our model
for the creation of culture-specific agents using only rituals is powerful enough to lead to the
perception of different cultures by the users. Thus, just by changing simple rituals in a set of
agents, one may be able to create a culture that is perceived by users as different.
For the second experiment, participants watched two cultures with the same rituals but with
an opposite dimension score: an individualistic high power distance culture and a collectivistic
high power distance culture. Like the first experiment, we asked participants similar questions
about the cultures. Yet, unlike the first experiment, we had many statistically significant
differences in the description of the characters, in terms of values and adjectives. In fact,
users did found the characters more individualistic or more collectivistic in accordance to
the parametrisation used. This is also a very encouraging result, which suggests that our
computational model of Hofstede’s dimensions can lead to the perception of the different
behavioural tendencies associated. However, 63% of participants attributed the differences
to the characters’ personalities instead of culture (again a statistically significant result).
Thus, comparing the results of the two experiments, we find that our dimensional compo-
nent is capable of differentiating cultures, yet the differences are not interpreted as cultural.
On the other hand, the rituals component is capable of leading to the perception of different
cultures, yet few differences are identified. Therefore, we believe that both these elements
are very important and should be considered when creating an agent architecture for creating
different cultures. Finally, the overall results we obtained suggest that the proposed model is
in fact capable of solving (at least partially) the problem we have raised in this dissertation.
7.1. FUTURE WORK 79
7.1 Future Work
This work can be continued following a few directions for future research regarding our cultural
model and its improvement:
• First, we believe the following evaluations would be interesting to perform with the model
as it is: (1) Repeating the experiments with most participants belonging to a culture dif-
ferent from Portugal, in order to investigate if the results would still rate the collectivistic
culture more favourably; (2) Repeat the rituals experiment but with the revised question-
naire so we could run statistical tests combining the data from both experiments; and (3)
Investigate the combined effect of rituals and the dimensions component in differentiating
extremes of the same dimension.
• Also, only two of the Hofstede’s dimensions were considered. We think the inclusion of
the other three is not trivial and would provide an interesting challenge.
• Currently, characters have only a single culture associated, which is based on findings
of national cultures. It would be interesting to extend the model in such a way that
it was possible for characters to also belong to many different subcultures, due to their
membership in certain ethnic, religious, and social groups.
• We also believe cultural conflict between characters would be a very interesting research
topic to explore. An interesting idea is characters having different predispositions on
cultural differences, based on the stages of the DMIS model we analysed in the Related
Work chapter.
• Finally, another promising topic for further research would be the role of culture in
affective computing. Our model already contemplates an emotional appraisal affected
by culture. However, there are many other relations between culture and emotions,
described in [42], that would be interesting to include, for example the notion of cultural
display rules (how should one act when experiencing certain emotions).
80 CHAPTER 7. CONCLUSION
Bibliography
[1] Dane Archer. Unspoken diversity: Cultural difference in gestures. Journal of Qualitative
Sociology, 20(1):79–105, 1997.
[2] M. B. Arnold and J. A. Gasson. Feelings and emotions as dynamic factors in personality
integration. New York: Ronald, 1954.
[3] R. Axelrod. The dissemination of culture. a model with local convergence and global
polarization. Journal of Conflict Resolution, 41(2):203–226, 1997.
[4] R. Aylett, S. Louchart, J. Dias, and A. Paiva. Fearnot! an experiment in emergent
narrative. In Springer, editor, Intelligent Virtual Agents, pages 305–316, 2005.
[5] R.F. Bales. Interaction Process Analysis. The University of Chicago Press, 1950.
[6] John H. M. Beattie. On Understanding Ritual. Basil Blackwell Press, 1970.
[7] Catherine Bell. Ritual: Perspectives and Dimensions. Oxford University Press, 1997.
[8] Milton Bennett. Towards ethnorelativism: A developmental model of intercultural sen-
sitivity. In R. M. Paige, editor, Cross-cultural orientation: New conceptualizations and
applications, pages 27 – 70. University Press of America, New York, 1986.
[9] Milton Bennett. Towards ethnorelativism: A developmental model of intercultural sensi-
tivity. In R. M. Paige, editor, Education for the intercultural experience, pages 21 – 71.
ME: Intercultural Press, Yarmouth, 1993.
[10] Milton Bennett. Overcoming the golden rule: Sympathy and empathy. In M.J. Bennett,
editor, Basic concepts of intercultural communication: A reader, pages 191 – 214. ME:
Intercultural Press, Yarmouth, 1998.
81
82 BIBLIOGRAPHY
[11] E. Bevacqua, A. Raouzaiou, C. Peters, G. Cardakis, K. Karpouzis C. Pelachaud, and
M. Mancini. Multimodal sensing, interpretation and copying of movements by a virtual
agent. In Perception and Interactive Technologies, pages 164 – 174. In PIT, 2006.
[12] Dorwin Cartwright and Alvin Zander. Group Dynamics: research and Theory. Harper
and Row, 1968.
[13] R. D’Andrade. Cultural meaning systems. In R. Schweder and R. Levine, editors, Culture
Theory. Essays on Mind, Self, and Emotion. Cambridge University Press, 1984.
[14] R. D’Andrade. Schemas and motivation. In R. D’Andrade and C. Strauss, editors, Human
motives and Cultural models, pages 23–44. Cambridge University Press, 1992.
[15] Eugene G. D’aquili. The myth-ritual complex: A biogenetic structural analysis. Journal
of Religion & Science, 18:247–269, 2005.
[16] J. Dias. Fearnot!: Creating emotional autonomous synthetic characters for empathic
interactions. Master’s thesis, Universidade Tecnica de Lisboa, Instituto Superior Tecnico,
Lisboa, 2005.
[17] J. Dias and A. Paiva. Feeling and reasoning: a computational model for emotional agents.
In Proceedings of 12th Portuguese Conference on Artificial Intelligence, EPIA 2005, pages
127–140. Springer, 2005.
[18] Joao Dias, Wan Ching Ho, Thurid Vogt, Nathalie Beeckman, Ana Paiva, and Elisabeth
Andre. I know what i did last summer: Autobiographic memory in synthetic characters.
In Ana Paiva, Rui Prada, and Rosalind W. Picard, editors, ACII, volume 4738 of Lecture
Notes in Computer Science, pages 606–617. Springer, 2007.
[19] Dietrich Dorner. The mathematics of emotions. In Prooceedings of the Fifth International
Conference on Cognitive Modeling, pages 75–79, Bamberg, Germany, 2003. Bamberg,
Germany.
[20] Mary Douglas. Natural Symbols: Explorations in Cosmology. New York: Random House,
New York, 2nd edition, 1973.
[21] R. Ewen. An introduction to theories of personality. Lawrence Erlbaum Associates, 1988.
BIBLIOGRAPHY 83
[22] Rui Figueiredo, Antonio Brisson, Ruth Aylett, and Ana Paiva. Emergent stories facili-
tated. In Ulrike Spierling and Nicolas Szilas, editors, ICIDS, volume 5334 of Lecture Notes
in Computer Science, pages 218–229. Springer, 2008.
[23] Michael P. Georgeff and Francois Felix Ingrand. Decision-making in an embedded reason-
ing system. In IJCAI, pages 972–978, 1989.
[24] G. Hofstede. Culture Consequences: Comparing Values, Behaviors, Intitutions, and Or-
ganizations Across Nations. Sage Publications, Thousand Oaks, 2001.
[25] G.J. Hofstede. Role playing with synthetic cultures: the evasive rules of the game. In Pro-
ceedings of the 9th International workshop of the IFIP, pages 49–56. Helsinky University
of Technology, SimLab Report no 10, 2005.
[26] G.J. Hofstede, P.B. Pedersen, and G. Hofstede. Exploring Culture - Exercises, Stories and
Synthetic Cultures. Intercultural Press, 2002.
[27] R. J. House, N. Wright, and R. N. Aditya. Cross-cultural research on organizational lead-
ership. cross cultural organizational behavior and psychology. In C. Earley and M. Eriz,
editors, New Perspectives on International Industrial/Organizational Psychology, pages
535 – 625. Pfeiffer, 1997.
[28] F. V. Jensen. Bayesian Networks and Decision Graphs. Springer, 2001.
[29] W. Lewis Johnson, Carole R. Beal, Anna Fowles-Winkler, Ursula Lauper, Stacy Marsella,
Shrikanth Narayanan, Dimitra Papachristou, and Hannes Hogni Vilhjalmsson. Tactical
language training system: An interim report. In James C. Lester, Rosa Maria Vicari, and
Fabio Paraguacu, editors, Intelligent Tutoring Systems, volume 3220 of Lecture Notes in
Computer Science, pages 336–345. Springer, 2004.
[30] W. Lewis Johnson, Hannes Hogni Vilhjalmsson, and Stacy Marsella. Serious games for
language learning: How much game, how much ai? In Chee-Kit Looi, Gordon I. McCalla,
Bert Bredeweg, and Joost Breuker, editors, AIED, volume 125 of Frontiers in Artificial
Intelligence and Applications, pages 306–313. IOS Press, 2005.
[31] I. Kant. Critique of Pure Reason, 2nd ed. New York: Macmillan Co., 2nd edition, 1787.
84 BIBLIOGRAPHY
[32] George A. Kelly. A theory of personality: The psychology of personal constructs. Norton,
1963.
[33] A.L. Kroeber and C. Kluckhohn. Culture: A Critical Review of Concepts and Definitions.
MA: Peabody Museum, Cambridge, 1952.
[34] Daniel Landis, Jannet M. Bennet, and Milton J. Bennet. Handbook of Intercultural Train-
ing. Sage Publications, Inc, 2003.
[35] Charles D. Laughlin. Ritual and symbolic function: A summary of biogenetic structural
theory. Journal of Ritual Studies, 4:16–39, 1990.
[36] Edmund Leach. The Political Systems of Highland Burma. London: Atholone Press, 1964.
[37] Mei Yii Lim, Joao Dias, Ruth Aylett, and Ana Paiva. Improving adaptiveness in au-
tonomous characters. In Helmut Prendinger, James C. Lester, and Mitsuru Ishizuka, ed-
itors, IVA, volume 5208 of Lecture Notes in Computer Science, pages 348–355. Springer,
2008.
[38] Steven Lukes. Political ritual and social integration. Journal of the British Sociological
Association, 9, 1975.
[39] Heidy Maldonado and Barbara Hayes-Roth. Toward cross-cultural believability in char-
acter design. In S. Payr and R. Trappl, editors, Agent Culture: Human-Agent Interaction
in a Multicultural World, pages 143 – 175. Lawrence Erlbaum Associates, London, 2004.
[40] H. Markus and S. Kitayama. Culture and the self: Implications for cognition, emotion,
and motivation. Psychological Review, 98:224–253, 1991.
[41] Joseph E. McGrath. Groups: Interaction and Performance. Prentice Hall, 1984.
[42] Batja Mesquita and Nico Frijda. Cultural variation in emotions: A review. Psychological
Bulletin, 112:179–204, 1992.
[43] M. Minsky. The psychology of computer vision., chapter A framework for representing
knowledge. New York: Mcgraw Hill, 1975.
[44] J. Muskus. Language study increases. yale daily news, nov. 21, 2003. Yale Daily News.
BIBLIOGRAPHY 85
[45] S. Nichols and S. Stich. Mindreading. An Integrated Account of Pretence, Self-Awareness,
and Understanding of Other Minds. Oxford University Press, 2003.
[46] A. Ortony, G. Clore, and A. Collins. The Cognitive Structure of Emotions. Cambridge
University Press, UK, 1988.
[47] Rui Prada. Teaming Up Human and Synthetic Characters. Phd thesis, Instituto Superior
Tecnico, 2005.
[48] Rui Prada and Ana Paiva. Beleivable groups of synthetic characters. In AAMAS’2005
- 4th International Joint Conference on Autonomous Agents and Multi Agent Systems,
Utrecht - Holland, 2005. ACM Press.
[49] David V. Pynadath and Stacy Marsella. Psychsim: Modeling theory of mind with decision-
theoretic agents. In IJCAI, pages 1181–1186, 2005.
[50] Matthias Rehm, Nikolaus Bee, Birgit Endrass, Michael Wissner, and Elisabeth Andre. Too
close for comfort?: Adapting to the user’s cultural background. In HCM ’07: Proceedings
of the international workshop on Human-centered multimedia, pages 85–94, New York,
NY, USA, 2007. ACM.
[51] Jeff Rickel and W. Lewis Johnson. Extending virtual humans to support team training
in virtual reality. pages 217–238, 2003.
[52] Ira Roseman and C. Smith. Appraisal Processes in Emotion: Theory, Methods, Research.
Oxford University Press, 2001.
[53] Robert Schank and R. Abelson. Scripts, Plans, Goals, and Understanding: An enquiry
into human knowledge structures. HillsDale, NJ: Earlbaum Assoc., 1977.
[54] Mei Si, Stacy Marsella, and David V. Pynadath. Thespian: Modeling socially normative
behavior in a decision-theoretic framework. In Jonathan Gratch, Michael Young, Ruth
Aylett, Daniel Ballin, and Patrick Olivier, editors, IVA, volume 4133 of Lecture Notes in
Computer Science, pages 369–382. Springer, 2006.
[55] S. Solomon, M. van Lent, M. Core, P. Carpenter, and M. Rosenberg. A language for
modeling cultural norms, biases and stereotypes for human behavior models. In BRIMS,
2008.
86 BIBLIOGRAPHY
[56] F. Thomas and O. Johnston. Disney Animation: The Illusion of Life. Abbeville Press,
New York, 1981.
[57] E.B. Tylor. Primitive culture: researches into the development of mythology, philosophy,
religion, art, and custom. New York Gordon Press, 1871.
[58] Valeri Valerio. Kingship and Sacrifice: Ritual and Society in Ancient Hawaii. Chicago:
University of Chicago Press, 1985.
[59] Margaret Visser. The Rituals of Dinner. Penguin, 1992.
Appendix A
Rituals Questionnaire
Questionnaire
Please, in order to answer the following questions, watch carefully the following video: First Video
Regarding the culture shown in the previous video, can you say what is more appropriate about it:
The eldest male should be the head of the household *
Yes No
Power and wealth are bad in this culture *
Yes No
It is important for leaders to make all the decisions *
Yes No
It is right for workers to openly disagree with their superiors *
Yes
No
They like to trust and
cooperate with other people *
Yes
No
It is important for them to
be independent * Yes
No
How would you characterise the group of people depicted in the video?
Please tick the value between -3 and 3 that fits best, with -3 representing the left adjective and +3 representing the right adjective.
If you think that neither the left nor the right adjective fits the group of people depicted
in the video, tick "0".
Approachable(-3) or Distant (3)? *
-3
-2
-1
0
1
2
3
Pleasant(-3) or Unpleasant
(3)? * -
3
-
2
-
1
0
1
2
3
Unfriendly(-3) or Friendly (3)? *
-3
-2
-1
0
1
2
3
Relaxed(-3) or Tense (3)? * -3
-2
-1
0
1
2
3
Compassionate(-3) or
Indifferent (3)? * -
3
-
2
-
1
0
1
2
3
Serious(-3) or Cheerful
(3)? * -
3
-
2
-
1
0
1
2
3
Warm(-3) or Cool (3)? * -3
-2
-1
0
1
2
3
Please, in order to answer the following questions, watch carefully this other video: Second Video
Regarding the culture shown in this second video, can you say what is more appropriate
about it:
The eldest male should be the head of the household *
Yes No
Power and wealth are bad in this culture *
Yes No
It is important for leaders to make all the decisions *
Yes No
It is right for workers to openly disagree with their
superiors * Yes
No
They like to trust and cooperate with other people *
Yes No
It is important for them to be independent *
Yes No
How would you characterise the group of people depicted in the second video?
Please tick the value between -3 and 3 that fits best, with -3 representing the left adjective and +3 representing the right adjective.
If you think that neither the left nor the right adjective fits the group of people depicted
in the video, tick "0".
Approachable(-3) or Distant (3)? *
-3
-2
-1
0
1
2
3
Pleasant(-3) or Unpleasant
(3)? * -
3
-
2
-
1
0
1
2
3
Unfriendly(-3) or Friendly (3)? *
-3
-2
-1
0
1
2
3
Relaxed(-3) or Tense (3)? * -3
-2
-1
0
1
2
3
Compassionate(-3) or
Indifferent (3)? * -
3
-
2
-
1
0
1
2
3
Serious(-3) or Cheerful
(3)? * -
3
-
2
-
1
0
1
2
3
Warm(-3) or Cool (3)? * -3
-2
-1
0
1
2
3
These next questions are about the two videos previously seen:
Did you notice differences in the two videos?
If yes, do you think most of those differences are related
to the culture or to the personality of the characters?
Finally, we would like to ask some information about yourself:
Gender? * Male
Female
Age? *
Nationality?
* = Input is required
Appendix B
Dimensions Questionnaire
Questionnaire
Please, in order to answer the following questions, watch carefully the following video:
First Video
Regarding the group of characters shown in the previous video, what do you feel about how they think and act?
Please select a value between -3 (strongly disagree) and 3 (strongly agree).
(Note: the idea is to give an opinion based only on your first impression of the characters)
The eldest male should be the head of the
household *
-
3
-
2
-
1
0
1
2
3
Power and wealth are bad *
-
3
-
2
-
1
0
1
2
3
They are concerned with everyone's well-
being? *
-
3
-
2
-
1
0
1
2
3
Personal achievements are very
important *
-
3
-
2
-
1
0
1
2
3
Direct confrontations should be avoided *
-
3
-
2
-
1
0
1
2
3
It is important for leaders to make all the
decisions *
-
3
-
2
-
1
0
1
2
3
It is right to openly disagree with
superiors *
-
3
-
2
-
1
0
1
2
3
They like to trust and cooperate with
other people *
-
3
-
2
-
1
0
1
2
3
It is important for them to be
independent *
-
3
-
2
-
1
0
1
2
3
How would you characterise the group of people depicted in the video?
Please select a value between -3 (left adjective) and 3 (right adjective).
If you think that neither the left nor the right adjective fits the group of people depicted in the video, tick "0".
Approachable(-3) or Distant (3)? *
-
3
-
2
-
1
0
1
2
3
Equal (-3) or Biased (3)? *
-
3
-
2
-
1
0
1
2
3
Independent(-3) or Sharing(3)? *
-
3
-
2
-
1
0
1
2
3
Equalitarian(-3) or Hierarchical (3)? *
-
3
-
2
-
1
0
1
2
3
Polite(-3) or Impolite(3)? *
-
3
-
2
-
1
0
1
2
3
Pleasant(-3) or Unpleasant (3)? *
-
3
-
2
-
1
0
1
2
3
Individualistic (-3) or Collectivistic (3)? *
-
3
-
2
-
1
0
1
2
3
Unfriendly(-3) or Friendly (3)? *
-
3
-
2
-
1
0
1
2
3
Relaxed(-3) or Tense (3)? *
-
3
-
2
-
1
0
1
2
3
Compassionate(-3) or Indifferent (3)? *
-
3
-
2
-
1
0
1
2
3
Serious(-3) or Cheerful (3)? *
-
3
-
2
-
1
0
1
2
3
Warm(-3) or Cool (3)? *
-
3
-
2
-
1
0
1
2
3
Now please watch carefully this other video:
Second Video
Regarding the group of characters shown in the second video, what do
you feel about how they think and act?
Please select a value between -3 (strongly disagree) and 3 (strongly agree).
(Note: the idea is to give an opinion based only on your first impression of
the characters)
The eldest male should be the head of the
household *
-
3
-
2
-
1
0
1
2
3
Power and wealth are bad *
-
3
-
2
-
1
0
1
2
3
They are concerned with everyone's well-
being? *
-
3
-
2
-
1
0
1
2
3
Personal achievements are very
important *
-
3
-
2
-
1
0
1
2
3
Direct confrontations should be avoided *
-
3
-
2
-
1
0
1
2
3
It is important for leaders to make all the
decisions *
-
3
-
2
-
1
0
1
2
3
It is right to openly disagree with
superiors *
-
3
-
2
-
1
0
1
2
3
They like to trust and cooperate with
other people *
-
3
-
2
-
1
0
1
2
3
It is important for them to be
independent *
-
3
-
2
-
1
0
1
2
3
How would you characterise the group of people depicted in the video?
Please select a value between -3 (left adjective) and 3 (right adjective).
If you think that neither the left nor the right adjective fits the group of people depicted in the video, tick "0".
Approachable(-3) or Distant (3)? *
-
3
-
2
-
1
0
1
2
3
Equal (-3) or Biased (3)? *
-
3
-
2
-
1
0
1
2
3
Independent(-3) or Sharing(3)? *
-
3
-
2
-
1
0
1
2
3
Equalitarian(-3) or Hierarchical (3)? *
-
3
-
2
-
1
0
1
2
3
Polite(-3) or Impolite(3)? *
-
3
-
2
-
1
0
1
2
3
Pleasant(-3) or Unpleasant (3)? *
-
3
-
2
-
1
0
1
2
3
Individualistic (-3) or Collectivistic (3)? *
-
3
-
2
-
1
0
1
2
3
Unfriendly(-3) or Friendly (3)? *
-
3
-
2
-
1
0
1
2
3
Relaxed(-3) or Tense (3)? *
-
3
-
2
-
1
0
1
2
3
Compassionate(-3) or Indifferent (3)? *
-
3
-
2
-
1
0
1
2
3
Serious(-3) or Cheerful (3)? *
-
3
-
2
-
1
0
1
2
3
Warm(-3) or Cool (3)? *
-
3
-
2
-
1
0
1
2
3
These next questions are about the two videos previously seen:
Did you notice differences in the two
videos?
If yes, do you think most of those
differences are related to the culture or to
the personality of the characters? Neithe
r
Cultur
e
Personalit
y