Content uploaded by Samuel Mascarenhas
Author content
All content in this area was uploaded by Samuel Mascarenhas on Feb 09, 2017
Content may be subject to copyright.
A Model of Social Dynamics for Social Intelligent Agents
Samuel Mascarenhas, Nuno Marques, Joana Campos and Ana Paiva
{smascarenhas,nuno.marques,joana.campos}@gaips.inesc-id.pt,ana.paiva@inesc-id.pt
INESC-ID and Instituto Superior T´
ecnico,
Universidade de Lisboa, Portugal
Abstract
In this article we describe a general cognitive model of
human social behavior that is meant to increase the so-
cial intelligence of autonomous intelligent agents in dif-
ferent contexts. Despite the remarkable improvements
that have been made on human-agent interaction, agents
still have a limited capacity to be aware of the social
reality that is present in the human mind and signifi-
cantly guides human behavior. The model discussed in
this paper is a step toward increasing that capacity sig-
nificantly. Two different case studies are described in
which the proposed model is used to better explain and
predict human behavior. The first case study is the well-
known Ultimatum game. The second one is a variation
of the “Game of Nines” played by children.
Introduction
One of the big dreams of Artificial Intelligence was, and still
is, the creation of machines, or agents, that perform tasks
and functions that require intelligence. Inspired by the sem-
inal work of Newell and Simon (Newell and Simon 1963),
for the past decades research in AI has grown significantly
trying to capture computationally the human’s ability for
high level, formal symbolic reasoning. In that context, intel-
ligence is mostly linked to rational, utility-based behaviour.
However, it is undeniable that aspects like emotions or social
behaviour are also required for achieving intelligence. So,
nowadays, many researchers in Artificial Intelligence have
been concerned with issues of believability and sociality in
the development of autonomous agents. These aspects turn
out to be particularly relevant when developing agents that
are meant to interact with humans, understand humans and
act in a natural manner with other agents (Reeves and Nass
1996).
Given that many different types of applications can ben-
efit from having agents that are socially-intelligent, their re-
search field is steadily growing. This includes applications
for education (Aylett et al. 2005; Paiva and Machado 1998;
Lester, Stone, and Stelling 1999; Lester et al. 1999), military
training (Johnson and Rickel 1997; Johnson, Vilhj´
almsson,
and Marsella 2005; Solomon et al. 2008), health (Bickmore
and Picard 2005; Johnson et al. 2003; Marsella, Johnson,
Copyright c
2013, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
and LaBore 2000), e-commerce (Cassell et al. 1999), enter-
tainment (Becker-Asano and Wachsmuth 2010; Prada and
Paiva 2009; Cassell et al. 2000; Tomlinson and Blumberg
2002; Rousseau and Hayes-Roth 1998), among others. In
the last decades, notable improvements have been made
in the creation of such agents, particularly in their emo-
tional capabilities(Dias and Paiva 2005; Gratch and Marsella
2004; Marsella, Gratch, and Petta 2010; Becker-Asano and
Wachsmuth 2010; Gebhard 2005; Rodrigues et al. 2009) and
conversational skills (Traum and Rickel 2002; Cassell et al.
1994; Gratch et al. 2007). Moreover, the inclusion of per-
sonality factors (Dias and Paiva 2005; McRorie et al. 2009;
Gebhard 2005; Egges, Kshirsagar, and Magnenat-Thalmann
2004; Rousseau and Hayes-Roth 1998) have facilitated the
creation of agents capable of expressing different individual
identities, enhancing their believability in a social context.
However, despite the improvements made, the capacity for
agents to simulate human social interaction is still strongly
limited as many challenges have yet to addressed. One such
challenge is the consideration of behavioural patterns that
occur at a social level of analysis. Even though we like to
view ourselves as strongly independent individuals, the re-
ality is that humans only managed to survive by living in
societal groups where resources had to be shared and labour
had to be divided. To solve such problems, among many oth-
ers, collaboration was essential. As a result, unwritten shared
rules and beliefs began to emerge in the earliest tribes. The
ones that were kept and passed on eventually became an in-
tegral part of the group’s shared knowledge.
Humans have the unique capability of creating a social
reality that is symbolic and exists only in the human mind.
Even though this reality cannot be directly observed, its in-
fluence in behaviour can be as strong as the physical real-
ity and often times stronger. This is evident when people go
against their survival instincts to defend an ideal or when
they intentionally starve in order to attain an idealized no-
tion of beauty. Such type of behaviour is difficult to explain
in agent models that only go so far as to include individual
factors, such as personality or biological drives. For agents
to interact with humans in a natural manner, they need to be
able to consider and be influenced by, at least partially, the
same type of social reality that drives human behaviour.
With the goal of better capturing this social reality that
exists in the human mind, in this article we describe the
39
Integrated Cognition
AAAI Technical Report FS-13-03
Social Importance Dynamics (SID) model, which is a cog-
nitive model of social behaviour that is grounded on the
status-power theory proposed by Kemper (Kemper 2011).
The model augments a typical BDI agent with a set of so-
cial dynamics that revolve around the concept of Social Im-
portance (SI). This concept is meant to represent the be-
havioural dimension of status as proposed in Kemper’s the-
ory. In this theory, the author divides human interaction in
two types: technical and relational. Technical activity cor-
responds to collaborative behaviour whose primary function
is to achieve an instrumental goal such as building a house
or winning a game of soccer. However, during technical ac-
tivity and outside of it, humans are constantly engaging in
relational behaviour. For instance, during a soccer match,
team mates will encourage one another and will celebrate
goals together. The aim of the SID model is to model the
motivation behind such type of behaviour.
The outline of this paper can be described as follows. In
the following section we describe the SID model, briefly dis-
cussing the theory in which the model is based and also pre-
senting how the model has an impact on the agent’s percep-
tion, deliberation and planning. Afterwards, we discuss the
application of the model in a well-known scenario, the Ul-
timatum Game, which has been extensively used to study
human decision-making. Finally, we explore the use of the
SID model in a more complex negotiation context, by us-
ing it to conduct a preliminary analysis of children playing a
variation of the “Game of Nines”, a game designed to evoke
conflict situations.
Social Importance Dynamics Model
According to the status-power theory (Kemper 2011), the
motivation behind all relational behaviour can ultimately be
explained in terms of two dimensions, status and power. Sta-
tus corresponds to our social standing in the eyes of another
person, which is then reflected on that person’s will to please
and act according to our interests instead of their own. Con-
versely, power corresponds to our ability to coerce another
person to comply with our wishes against their will.
The two dimensions have quite different dynamics associ-
ated to them. For instance, power can be obtained by holding
a weapon against a person whereas status originates from
positive social interaction. As our primary concern is to im-
prove human-agent interaction, we started by only address-
ing the notion of status in the SID model. Thus, the model
makes the assumption that neither agents nor users will at-
tempt to coerce or manipulate others. As such, the inclu-
sion of power is left as future work. It is also important to
distinguish the concept of status as proposed in Kemper’s
theory from the common use of the word to refer to hierar-
chical roles in a social group such as “boss”, “teacher”, or
“mother”. To explain the distinction, consider the following
examples. A boss can have a very high or very low status
depending on how well he is respected and admired by his
employees but, in both cases, his power is high given that he
can fire his staff if they fail to meet his expectations. In the
case of a typical mother-son relationship, although power-
less, the son has a very high status, given that the mother is
willing to do everything she can for her son. To avoid confu-
sion between the two meanings, our proposed model refers
to its implementation of status as social importance (SI).
The SI concept is modelled as a numeric variable that
is associated to a relationship between two agents. More
precisely, SI(A, B ), is a scalar that represents how much
is agent A willing to act in the interest of agent B, which
might be different than what agent B is willing to do for A,
i.e. SI (A, B)is not the same as SI(B , A). But how is it
determined? The amount of SI that a person attributes to
another is determined by many different sources including
friendship, family ties, group membership, behavioral con-
duct, admirable qualities, trust, reputation, among others. In
this sense, SI constitutes an integrative measure of social
factors in a similar manner that the concept of utility rep-
resents an integrative measure of motivational factors. The
main advantage of having this broad social concept is that
it allows us to abstract from the particularities of the differ-
ent relational factors and model a common underlying pro-
cess. For instance, the relationship between father and son
is clearly distinct from a lovers relationship in terms of ap-
propriate behaviours. However, in both cases, the extent that
one is willing to act in the interest of the other is similarly
high, which is what the concept of SI represents.
According to the status-power theory (Kemper 2011), our
actions towards others can be perceived as either claims
or conferrals, from a relational standpoint. A claim occurs
whenever someone makes a request to another. The bigger
the request, the higher the claim. After a claim is performed,
a conferral act is expected in response, with different actions
conferring different amounts of SI. For illustration purposes,
consider a professor that is about to give an invited talk on a
conference but is not able to find the room. The professor can
approach a random person in the lobby and ask for directions
and this would be perceived as a relatively low claim. Then,
based on the SI that the other person attributes to the profes-
sor, a decision is made on how much will the person confer.
The following three actions, exemplify three possible con-
ferrals that vary on the amount of SI conferred (from lowest
to highest): (1) come up with an excuse to avoid having to
explain the directions, (2) take the time to explain the direc-
tions or (3) personally accompany the professor to the room.
A person who highly admires the professor will be inclined
to choose the third option. Oppositely, a person who has very
little respect for the professor is more likely to choose the
first option instead.
Figure 1: General diagram of the SID model
40
Based on the aforementioned notions, the SID
model is based on the following three components:
(1) SI AttributionRules, (2) SIC onferrals, and (3)
SI Claims. As illustrated in Figure 1, each of these compo-
nents has an impact on a different process of a general BDI
architecture. The SI AttributionRules are used to deter-
mine how much SI should be attributed to another agent.
These are formally defined as a <T,A,V>tuple where: T
corresponds to the target of the rule, A is a list of conditions
that specify when the rule is activated, and V specifies the
amount of SI the target of the rule gains/loses while the rule
is active. The impact of these rules on the agent’s perception
corresponds to having the agent actively checking which
rules can be activated for every other agents, every time
they update their beliefs. Additionally, agents also infer how
much SI do they have in the perspective of others, assuming
that these others adopt an identical set of rules.
The SI Conf errals have an impact on the agent’s delib-
eration by automatically generating social goals to confer
status to others, which according to (Kemper 2011), corre-
sponds to an intrinsic human motivation that explains why
people care so much about greeting rituals and other activi-
ties that achieve no instrumental function outside of the so-
cial world. Formally, an SI Conf erral is defined by the tu-
ple <T,C,A,V>where: T is the target agent to whom the
conferral is directed, C is a set of activation conditions that
dictate the context in which the conferral is relevant, A is the
name of the conferral action that needs to be performed by
the agent and finally V defines the amount conferred by the
conferral.
If there is a target agent that has the same or more SI than
the amount conferred by an SI Conf erral and the context
conditions of the conferral are all verified, a goal to execute
the conferral act is automatically generated. The utility of
this goal is linearly proportional to the amount of the SI the
act confers. Consequently, if an agent has two or more con-
ferral goals activated that have the same target, the conferral
selected will be the one that confers the most, without ex-
ceeding the SI of the target.
The last component of the SID model is the SI Claims.
These are represented as a <A,V>pair where: A is the the
name of the action that symbolizes a claim and V is the
amount of SI that is claimed by the action. They affect the
planning process in the following way. While the agent is
planning a course of actions to achieve its intention, it will
check if any action of the current plan corresponds to an
SI Claim. When this happens, the agent will check its as-
sumption of how much SI it has on the perspective of the
target agent. If the agent infers that its SI is not enough, then
the action is removed and the agent searches for an alterna-
tive plan that also achieves the same goal. This implies that
agents will never try to claim more SI than what they believe
they have. This is a simplification we plan to address in fu-
ture work, as sometimes this form of behaviour is performed
by humans in an intentional manner.
To clarify how these components can be used to model
social behaviour, the previous situation of the lost professor
will be used as an example. This particular situation can be
modelled with the following SIAttributionRules:
•<T = x, isPerson(x) = True, V = +30>
•<T = x, isAdmired(x) = True, V = +10>
Only one claim is needed, namely:
•<A = ask-directions, V = 30>.
Finally, in response to this claim, consider that the possi-
ble conferrals are:
•<T = x, C = performed(x,ask-directions), A = give-
excuse,V = 20>
•<T = x, C = performed(x,ask-directions), A = give-
directions,V = 30>
•<T = x, C = performed(x,ask-directions), A =
accompany-to-destination, V = 40>
When the professor and the random person perceive each
other, they will attribute some amount of SI to one another.
Based on the first rule, the professor will attribute to the per-
son an SI of 30. However, the random person will attribute
to the professor an SI of 40 or 30, depending if that person
admires the professor or not (second attribution rule). The
professor, aware that he is lost, forms an intention to obtain
directions. When planning to achieve this intention, he plans
to perform the action of asking directions to the random per-
son. Because asking directions is a claim of 30, the profes-
sor will consider that this is a socially appropriate action to
perform to the person and does so. In response to the claim
made by the professor, the person activates all the three pos-
sible conferrals and selects the highest one that does not ex-
ceed the SI attributed to the professor. This means that if
the person admires the professor, then the third conferral is
chosen and the person will accompany the professor. If not,
then the person will select the second conferral and will just
explain the directions.
Ultimatum Game
Originally, the SID model was used to facilitate the creation
of agents with different cultural biases in their behaviour.
This involved the implementation of the model in FAtiMA
(Dias, Mascarenhas, and Paiva 2011), an existing BDI archi-
tecture for embodied agents (Dias, Mascarenhas, and Paiva
2011). The resulting architecture was then applied to cre-
ate an intercultural training tool where users can socially in-
teract with agents that have different cultural profiles (Mas-
carenhas et al. 2013).
Although the SID model was originally meant to augment
the social intelligence of a typical BDI agent, we are also in-
terested in exploring its use in more formalized scenarios
such as the Ultimatum Game. The motivation for doing so
comes from the fact that this scenario has been studied ex-
tensively and it is a good example of how human behaviour
deviates from optimal game-theoretic predictions. As such,
we are interested in determining if the social dynamics pro-
posed in the SID model are able to more closely reflect how
humans behave when playing this game.
In the Ultimatum Game, first developed by (G¨
uth, Schmit-
tberger, and Schwarze 1982), two players must divide a
given sum of money between them. One of the players, the
41
Allocator, first chooses how that amount should be be split.
The other player, the Recipient, can then either accept or re-
ject the offer. If he or she accepts, then each player receives
the agreed amount. Otherwise, both get nothing. In its stan-
dard form, the game is one shot.
Despite its very simple structure, the Ultimatum Game
can be found in many every day situations (eg. bargaining),
making it a specially attractive tool in studies of social de-
cision making (Handgraaf, Van Dijk, and De Cremer 2003).
From a game theoretic perspective, the Allocator should pro-
pose the minimum possible amount, which would always be
accepted by the Recipient (getting something is better than
getting nothing). However, empirical data shows that the fair
offer (a 50/50 split) is the most common outcome (Hoff-
man, McCabe, and Smith 1996). Thus, it seems as people
take other factors into consideration besides their personal
expected return when playing the Ultimatum Game.
The aforementioned factors appear to be of socio-cultural
nature. For instance, (Henrich et al. 2001) found that the
mean Allocator offers and rejection rates may depend on the
nationality of the participants, and (Bolton and Zwick 1995)
showed that a lower social distance between the participants
result in higher offers and higher acceptance rates. These
findings are reinforced by neuroscience research linking the
activation of specific areas of the brain with rewards of social
nature, such as equality and reciprocity (Sanfey et al. 2003;
Sanfey 2007).
In the context of our work, this effects may be represented
by employing the SID model in an Ultimatum Game Sce-
nario. An Allocator agent (A) determines his offer, p, using
the amount of SI he attributes to the Recipient, such that:
p=SI (A, R)∗k1
In the case of a Recipient agent (R), the decision whether
to accept or reject depends on the agent’s expectation over
the offer he will receive. This expectation is a scalar e∈
[0,1], which is linearly proportional to the Receiver’s esti-
mation of the SI attributed to him by the Allocator, repre-
sented by SIR(A, R). Thus:
e=SIR(A, R)∗k2
This expectation can then be used to determine the Recip-
ient’s response:
f(p) = accept if p−e>ε
reject otherwise
where εis the error margin associated with the expectation.
We are preparing a user study where human opponents
play against agents, so we can gather data on how people
perceive the emerging behaviour. This will imply first ma-
nipulating the SI attributed to the agent, for example via
in-group/out-group membership information. Then we can
check if the agent predictions match the way the human
plays, and thus if it reacts realistically.
In addition, the social dynamics of the game will also be
explored by varying the amount of SI and, consequently, the
amounts, as a result of making unfair offers or rejecting fair
offers. We will analyse at what level does the “social fair-
ness” factor, represented by the SI, interacts with the self-
interest drive, and balance their effects in order to achieve
more human-like behaviours.
We should note, however, that other factors, such as Kem-
per’s notion of power and a self-interest drive are still absent
and their inclusion is seen as future work. However, this pre-
liminary approach will likely help us in refining the model.
Game of Nines
Although the repeated ultimatum game is a good game to ex-
plore the impact of social relations in decision making, we
also wanted to investigate how social importance is linked
to the way we negotiate. To do that, and thus investigating
the types of interactions that emerge and how the social im-
portance is linked to negotiation, we have designed an ex-
periment with the “Game of Nines”. The “Game of Nines”
is a mixed-motive bargaining game and it was firstly used
by Kelley et al. (Kelley, Beckman, and Fischer 1967). The
two-person game requires that the players start a negotiation
process to divide a joint reward between themselves.
Each player holds cards from one to eight in their hands.
At each trial, the negotiators must agree on the cards they
play such that their sum is nine or less. Furthermore, for each
bargaining problem each negotiator is privately assigned to a
minimum necessary share (MNS). The negotiator must play
a card above the MNS if he wants the agreement to be prof-
itable to him (e.g. if a player has a MNS equals to 3 and plays
a 6 he will get 3 as a reward). If no agreement is reached both
players get zero. A person does not have direct knowledge
of the other’s MNS value so the negotiation occurs under in-
complete information. Adding to that, the negotiation must
occur in a limited amount of time.
These conditions create a mixed-motive relationship in
the sense that it is of their mutual interest to agree upon a
division that gives them a share larger than their MNS in
the minimum use of time. On the other hand, it is of each
person’s individual interest to guarantee to have the largest
share in the division at each trial.
This bargaining game creates an interesting setting, where
the negotiators face dilemmas concerning their goals and
forms of communication. We used this game to investigate
forms of social conflict and the dynamics behind the social
interactions (thus linked to the SID model we have been
investigating). In particular, we wanted to understand how
children perceive situations of disadvantage and conflicting
interests and which strategies they employ in different rela-
tional contexts.
We have designed an experiment, involving 11 pairs of
children aged 10 to 12 years-old. Opt-out consent forms
were provided to all parents or guardians of those children.
All games were video and audio recorded.
Method
The basic paradigm of the “Game of Nines” was used in
our experiment, a constant-sum game with incomplete in-
formation. Yet, as the experiment subjects were children we
42
eliminated the time constraint1.
The experiment consisted in 5 rounds. At the beginning
of each round the subjects took from an envelop a number
that represents their minimum necessary share (MNS). The
players were told to never show that number to the opponent
during the trial and never agree on a value below that num-
ber. At the end of the round the participants have to show
their MNS value to the other.
On each bargaining trial the players have to jointly agree
on a possible contract. Each contract corresponds to a card
that is going to be played by player A and a card played by
player B, so that their sum does not exceed 9. For example,
if player A plays the card 7 a possible contract is player B to
play card 2. The interests of the parties are always directly
opposed. Besides the cards ranging from one to eight that
each player holds, they also have a card that allows them
to give up if they feel they are not able to achieve a viable
agreement.
The MNS values throughout the rounds are in table 1. The
combinations of values at each trial, although different, they
were chosen to describe the same situation, i.e. the distri-
bution of viable contracts is the same throughout the trials.
An exception occurs at the 4th round, which given the MNS
values there are not any mutually profitable agreement.
Player A Player B
Round 1 2 2
Round 2 1 3
Round 3 3 1
Round 4 5 5
Round 5 2 2
Table 1: MNS values for each player per round
To motivate the participants to do well, we told the players
that the winner of the game would win a prize. The winner
is the player who accumulates more points throughout the
rounds. This is a constant-sum game, so the points at round
3 are the sum of points gained at the previous rounds. The
number of points a players gets at each round is determined
by the card played minus his minimum at that round. After
the explanation of the rules, the participants were “walked
through” two rounds of the game to learn its mechanics. Fol-
lowing that they were left alone to play the game.
Analysis of Negotiation Behaviours
This experimental setting incorporates motivational, cogni-
tive, perceptual and decision making processes. As a possi-
ble way to understand its dynamics we used the Social Im-
portance Dynamics Model to examine how these processes
unfold.
For illustration purposes, consider the following interac-
tion between two girls (see Figure 2) during the first round of
the game (the MNS values are in Table 1). Between brackets
is the analysis of the interaction according to the SID model.
The acronyms used in the excerpt are described in Table 2.
1We verified in pilot sessions that this factor was making them
not to pay attention to what was happening in the game.
CLM-H High Claim
CLM-L Low Claim
CONF-L Conferral lower than expected by the
claim
CONF-EXP Conferral expected in response to the
claim
CONF-SP Spontaneous conferral with no explicit
claim
Table 2: Acronyms used in the analysis with the SID model.
Player B: So...
Player A: OK. You play first. (CONF-SP)
Player B: I play a ....<hesitation>.
But first we have to see our
minimums don’t we?
A: First we have to take the
minimum.
< They take the envelop for that round
and save their card>
B: I play the card "6". (CLM-H)
Do you agree or not?
A: I agree. (CONF-EXP)
B: Clan I play the card? (CLM-H)
<She places the card on the table>
A: Yes. (CONF-EXP)
A: I play... <hesitation>
B: <Coff>. huh?...
A: <Look at the cards>
B: Are you going to play?
You play the card "3". Yes? (CLM-H)
A: Wait. Wait. Wait. (CONF-L)
B: I’m confused....
A: Wait. I play the card 6. (CLM-H)
B: What? (CONF-L)
A: I play 6. (CLM-H)
B: Huh? (CONF-L)
A: I play 6. (CLM-H)
B: But that doesn’t work. 6+6 is 12.
(CONF-L)
A: That’s right...
B: The maximum is 9.
Now I’m a little bit confused.
A: OK. I play 4. (CLM-L)
B: That doesn’t work either. (CONF-L)
A: 7,8,9.
B: It can’t be. Gives 10.
A: I play 3. (CONF-EXP)
A: <laugh> (CONF-SP)
B: <laugh> (CONF-SP)
This example suggests that Player A attributes more SI to
Player B than the opposite. The first sign of this occurs at the
start of the interaction where Player A proposes that Player
B should be the first to play. Then, Player B makes a really
high claim (she proposes to play 6) and Player A accepts
it straight away. But at some point in the interaction, while
43
Player A is deciding the card she will play, she realizes that
Player B is being a bit greedy. But instead of directly ask-
ing Player B to play a lower card, she says that she will also
play 6, possibly expecting that player B goes back in her
initial proposal. Player B quickly expresses her surprise and
confusion regarding Player A’s attempt. She then continu-
ously rejects all the offers that Player A makes, until, finally,
Player A accepts to play 3. Interestingly, they both laugh at
the end. This exchange of conferrals possibly serves the pur-
pose of re-establishing any loss of SI that occurred during
the round.
It should be emphasized that this is a tentative explanation
for the observed phenomena. Other factors may be influenc-
ing the player’s behaviour, such as a limited understanding
of game rules. A more in-depth analysis is thus warranted
in order to isolate the effects of Social Importance in this
negotiation game.
Figure 2: Two girls playing the Game of Nines
We are now analyzing and annotating the interactions of
all the eleven dyads using the SID model, not only to inform
us about the social dynamics that unfolds during the games,
but also, as a way to inform the model itself and thus, further
refine it.
Conclusion
In this paper we argued about the importance of increas-
ing the social intelligence of autonomous agents by mak-
ing them more aware of the relational aspects of human
behaviour. More than physically instrumental, our actions
carry out symbolic meanings and their performance signifies
our relation with others. In this paper, we described a cogni-
tive model of social behaviour that takes into account these
relational meanings and endows agents with a set of social
dynamics concerning them. More precisely, the proposed
SID model affects the perception of agents by making them
aware of their relational standing with other agents, which
is treated as a numeric variable, named SI. The model af-
fects deliberation by automatically creating conferral goals
to convey how much SI do agents attribute to others. The
model also affects planning by making agents avoiding ac-
tions that would claim more SI than what other agents at-
tribute to them.
For the sake of exploring how the proposed SID model
can be used to simulate human decision-making, we dis-
cussed its implementation in the Ultimatum Game. This is a
well-researched game that is played by two players in which
one of them makes a monetary offer to the other and the
other can either accept or reject the offer made. Previous
studies have shown that people are significantly influenced
by socio-cultural factors when playing this game and we are
attempting to represent such factors with the SID model. We
plan to conduct different manipulations on the SI that either
the player or the agent attributes to one another and observe
how it affects their strategy. We are also interested in seeing
how the SID model can be useful to model more complex
games, that involve active negotiation between the players.
For this purpose, we have recorded 11 pairs of children play-
ing a variation of the “Game of Nines”, a game in which two
players must bargain a division between nine points with one
another across multiple rounds. We are currently testing the
predictive power of the SID model by using it to analyse
and describe the observed interactions between the children
playing the game as it was exemplified in this article.
Acknowledgements
This work was partially supported by national funds through
FCT - Fundac¸˜
ao para a Ciˆ
encia e a Tecnologia, under project
PEst-OE/EEI/LA0021/2013. It was also partially sup-
ported by two FCT scholarships (SFRH BD /62174/2009),
(SFRH/BD/75342/2010) and by the European Community
(EC), being partially funded by the ECUTE project (ICT-5-
4.2 257666), the EMOTE project (FP7 ICT-317923) and the
SIREN project (FP7 ICT 258453). The authors are solely
responsible for the content of this publication. It does not
represent the opinion of the EC or the FCT, which are not
responsible for any use that might be made of data appear-
ing therein.
References
Aylett, R.; Louchart, S.; Dias, J.; Paiva, A.; and Vala, M.
2005. Fearnot! - an experiment in emergent narrative.
In Panayiotopoulos, T.; Gratch, J.; Aylett, R.; Ballin, D.;
Olivier, P.; and Rist, T., eds., Intelligent Virtual Agents, vol-
ume 3661 of Lecture Notes in Computer Science. Springer
Berlin Heidelberg. 305–316.
Becker-Asano, C., and Wachsmuth, I. 2010. Affective com-
puting with primary and secondary emotions in a virtual hu-
man. Autonomous Agents and Multi-Agent Systems 20:32–
49.
Bickmore, T. W., and Picard, R. W. 2005. Establishing and
maintaining long-term human-computer relationships. ACM
Transactions on Computer-Human Interaction 12(2):293–
327.
Bolton, G. E., and Zwick, R. 1995. Anonymity versus pun-
ishment in ultimatum bargaining. Games and Economic Be-
havior 10(1):95–121.
Cassell, J.; Pelachaud, C.; Badler, N.; Steedman, M.;
Achorn, B.; Becket, T.; Douville, B.; Prevost, S.; and Stone,
44
M. 1994. Animated conversation: rule-based generation
of facial expression, gesture & spoken intonation for mul-
tiple conversational agents. In Proceedings of the 21st an-
nual conference on Computer graphics and interactive tech-
niques, SIGGRAPH ’94, 413–420. New York, NY, USA:
ACM.
Cassell, J.; Bickmore, T.; Billinghurst, M.; Campbell, L.;
Chang, K.; Vilhj´
almsson, H.; and Yan, H. 1999. Embod-
iment in conversational interfaces: Rea. In Proceedings of
the SIGCHI conference on Human Factors in Computing
Systems, CHI ’99, 520–527. New York, NY, USA: ACM.
Cassell, J.; Ananny, M.; Basu, A.; Bickmore, T.; Chong, P.;
Mellis, D.; Ryokai, K.; Smith, J.; Vilhj´
almsson, H.; and Yan,
H. 2000. Shared reality: physical collaboration with a virtual
peer. In CHI ’00 Extended Abstracts on Human Factors in
Computing Systems, CHI EA ’00, 259–260. New York, NY,
USA: ACM.
Dias, J., and Paiva, A. 2005. Feeling and reasoning: a com-
putational model for emotional agents. In Proceedings of
12th Portuguese Conference on Artificial Intelligence, EPIA
2005, 127–140. Springer.
Dias, J.; Mascarenhas, S.; and Paiva, A. 2011. Fatima modu-
lar: Towards and agent architecture with a generic appraisal
framework. In Proceedings of the Workshop on Standards
in Emotion Modelling. Leiden: Lorentz Center.
Egges, A.; Kshirsagar, S.; and Magnenat-Thalmann, N.
2004. Generic personality and emotion simulation for con-
versational agents. Computer Animation and Virtual Worlds
15(1):1–13.
Gebhard, P. 2005. Alma: a layered model of affect. In
Proceedings of the fourth international joint conference on
Autonomous agents and multiagent systems, AAMAS ’05,
29–36. New York, NY, USA: ACM.
Gratch, J., and Marsella, S. 2004. A domain-independent
framework for modeling emotion. Cognitive Systems Re-
search 5(4):269–306.
Gratch, J.; Wang, N.; Gerten, J.; Fast, E.; and Duffy, R.
2007. Creating rapport with virtual agents. Intelligent Vir-
tual Agents 125–138.
G¨
uth, W.; Schmittberger, R.; and Schwarze, B. 1982. An
experimental analysis of ultimatum bargaining. Journal of
Economic Behavior & Organization 3(4):367–388.
Handgraaf, M. J.; Van Dijk, E.; and De Cremer, D. 2003. So-
cial utility in ultimatum bargaining. Social Justice Research
16(3):263–283.
Henrich, J.; Boyd, R.; Bowles, S.; Camerer, C.; Fehr, E.;
Gintis, H.; and McElreath, R. 2001. In search of homo eco-
nomicus: behavioral experiments in 15 small-scale societies.
The American Economic Review 91(2):73–78.
Hoffman, E.; McCabe, K.; and Smith, V. 1996. On expecta-
tions and the monetary stakes in ultimatum games. Interna-
tional Journal of Game Theory 25:289–301.
Johnson, W. L., and Rickel, J. 1997. Steve: An animated
pedagogical agent for procedural training in virtual environ-
ments. ACM SIGART Bulletin 8(1-4):16–21.
Johnson, W. L.; Shaw, E.; Marshall, A.; and LaBore, C.
2003. Evolution of user interaction: the case of agent adele.
In Proceedings of the 8th international conference on Intelli-
gent user interfaces, IUI ’03, 93–100. New York, NY, USA:
ACM.
Johnson, W. L.; Vilhj´
almsson, H. H.; and Marsella, S. 2005.
Serious games for language learning: How much game, how
much ai? In Looi, C.-K.; McCalla, G. I.; Bredeweg, B.; and
Breuker, J., eds., AIED, volume 125 of Frontiers in Artificial
Intelligence and Applications, 306–313. IOS Press.
Kelley, H. H.; Beckman, L. L.; and Fischer, C. S. 1967. Ne-
gotiating the division of a reward under incomplete informa-
tion. Journal of Experimental Social Psychology 3(4):361 –
398.
Kemper, T. 2011. Status, power and ritual interaction: a
relational reading of Durkheim, Goffman, and Collins. Eng-
land: Ashgate Publishing Limited.
Lester, J. C.; Voerman, J. L.; Towns, S. G.; and Callaway,
C. B. 1999. Deictic believability: Coordinating gesture, lo-
comotion, and speech in lifelike pedagogical agents. Applied
Artificial Intelligence 13:383–414.
Lester, J. C.; Stone, B. A.; and Stelling, G. D. 1999. Lifelike
pedagogical agents for mixed-initiative problem solving in
constructivist learning environments. User Modeling and
User-Adapted Interaction 9(1-2):1–44.
Marsella, S.; Gratch, J.; and Petta, P. 2010. Computational
Models of Emotion. Oxford University Press.
Marsella, S. C.; Johnson, W. L.; and LaBore, C. 2000. In-
teractive pedagogical drama. In Proceedings of the fourth
international conference on Autonomous agents, AGENTS
’00, 301–308. New York, NY, USA: ACM.
Mascarenhas, S.; Prada, R.; Paiva, A.; and Hofstede, G. J.
2013. Social importance dynamics: A model for culturally-
adaptive agents. In Aylett, R.; Krenn, B.; Pelachaud, C.;
and Shimodaira, H., eds., Intelligent Virtual Agents, vol-
ume 8108 of Lecture Notes in Computer Science, 325–338.
Springer.
McRorie, M.; Sneddon, I.; Sevin, E.; Bevacqua, E.; and
Pelachaud, C. 2009. A model of personality and emotional
traits. In Ruttkay, Z.; Kipp, M.; Nijholt, A.; and Vilhjalms-
son, H., eds., Intelligent Virtual Agents, volume 5773 of Lec-
ture Notes in Computer Science. Springer Berlin Heidelberg.
27–33.
Newell, A., and Simon, H. 1963. GPS: a program that sim-
ulates human thought. In Feigenbaum, E. A., and Feldman,
J., eds., Computers and Thought. McGraw-Hill. 279–293.
Paiva, A., and Machado, I. 1998. Vincent, an autonomous
pedagogical agent for on-the-job training. In Goettl, B.;
Halff, H.; Redfield, C.; and Shute, V., eds., Intelligent Tu-
toring Systems, volume 1452 of Lecture Notes in Computer
Science. Springer Berlin Heidelberg. 584–593.
Prada, R., and Paiva, A. 2009. Teaming up humans
with autonomous synthetic characters. Artificial Intelligence
173(1):80–103.
Reeves, B., and Nass, C. 1996. The Media Equation - How
45
People Treat Computers, Television, and New Media Like
Real People and Places. Cambridge University Press.
Rodrigues, S. H.; Mascarenhas, S. F.; Dias, J.; and Paiva, A.
2009. i can feel it too!: Emergent empathic reactions be-
tween synthetic characters. In Proc. of the 3rd International
Conference on Affective Computing and Intelligent Interac-
tion (ACII) 2009, 1–7. IEEE.
Rousseau, D., and Hayes-Roth, B. 1998. A social-
psychological model for synthetic actors. In Proceedings of
the second international conference on Autonomous agents,
165–172. ACM.
Sanfey, A. G.; Rilling, J. K.; Aronson, J. A.; Nystrom,
L. E.; and Cohen, J. D. 2003. The neural basis of eco-
nomic decision-making in the ultimatum game. Science
300(5626):1755–1758.
Sanfey, A. G. 2007. Social decision-making: insights from
game theory and neuroscience. Science 318(5850):598–602.
Solomon, S.; van Lent, M.; Core, M.; Carpenter, P.; and
Rosenberg, M. 2008. A language for modeling cultural
norms, biases and stereotypes for human behavior models.
In BRIMS.
Tomlinson, B., and Blumberg, B. 2002. Social synthetic
characters. Computer Graphics 26(2).
Traum, D., and Rickel, J. 2002. Embodied agents for multi-
party dialogue in immersive virtual worlds. In Proceedings
of the first international joint conference on Autonomous
agents and multiagent systems: part 2, AAMAS ’02, 766–
773. New York, NY, USA: ACM.
46
