How Women Think Robots Perceive Them – as if Robots were Men

Matthijs A. Pontier and Johan F. Hoorn

Center for Advanced Media Research Amsterdam (CAMeRA@VU) / The Network Institute, VU University,

De Boelelaan 1081, 1081HV, Amsterdam, The Netherlands,

Keywords: Cognitive Modeling, Emotion Modeling, Human-computer Interaction, Turing Test, Virtual Agents.

Abstract: In previous studies, we developed an empirical account of user engagement with software agents. We

formalized this model, tested it for internal consistency, and implemented it into a series of software agents

to have them build up an affective relationship with their users. In addition, we equipped the agents with a

module for affective decision-making, as well as the capability to generate a series of emotions (e.g., joy

and anger). As follow-up of a successful pilot study with real users, the current paper employs a non-naïve

version of a Turing Test to compare an agent’s affective performance with that of a human. We compared

the performance of an agent equipped with our cognitive model to the performance of a human that

controlled the agent in a Wizard of Oz condition during a speed-dating experiment in which participants

were told they were dealing with a robot in both conditions. Participants did not detect any differences

between the two conditions in the emotions the agent experienced and in the way he supposedly perceived

the participants. As is, our model can be used for designing believable virtual agents or humanoid robots on

the surface level of emotion expression.

1 INTRODUCTION

1.1 Background

There is a growing interest in developing embodied

agents and robots. They can make games more

interesting, accommodate those who are lonely,

provide health advice, make online instructions

livelier, and can be useful for coaching, counselling,

and self-help therapy. In extreme circumstances,

robots can also be the better self of human operators

in executing dangerous tasks.

For a long time, agents and social robots were

mainly developed from a technical point of view but

we now know it is not a matter of technology alone.

Theories and models of human life are also important

to explain communication rules, social interaction and

perception, or the appraisal of certain social situations.

In media psychology, mediated interpersonal

communication and human-computer interaction,

emotions play a salient role and cover an important

area of research (Konijn and Van Vugt, 2008).

The idea of affective computing (Picard, 1997) is

that computers ‘have’ emotions, and detect and

understand user emotions to respond appropriately to

the user. Virtual agents who show emotions may

increase the user’s likeability of a system. The

positive effects of showing empathetic emotions are

repeatedly demonstrated in human-human

communication (e.g., Konijn and Van Vugt, 2008)

and are even seen as one of the functions of emotional

display. Such positive effects may also hold when

communicating with a virtual agent. Users may feel

emotionally attached to virtual agents who portray

emotions, and interacting with such “emotional”

embodied computer systems may positively influence

their perceptions of humanness, trustworthiness, and

believability. User frustration may be reduced if

computers consider the user’s emotions (Konijn and

Van Vugt, 2008). A study by Brave et al. (2005)

showed that virtual agents in a blackjack computer

game who showed empathic emotions were rated

more positively, received greater likeability and

trustworthiness, and were perceived with greater

caring and support capabilities than virtual agents not

showing empathy.

Compared to human affective complexity,

contemporary affective behavior of software agents

and robots is still quite simple. In anticipation of

emotionally more productive interactions between

user and agent, we looked at various models of

human affect-generation and affect-regulation, to see

how affective agent behavior can be improved.

496

A. Pontier M. and F. Hoorn J..

How Women Think Robots Perceive Them – as if Robots were Men.

DOI: 10.5220/0004253504960504

In Proceedings of the 5th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2013), pages 496-504

ISBN: 978-989-8565-39-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

1.2 From Theories to Computation

Previous work described how certain dimensions of

synthetic character design were perceived by users

and how they responded to them (Van Vugt et al.,

2009). A series of user studies into human-agent

interaction resulted into an empirically validated

framework called Interactively Perceiving and

Experiencing Fictional Characters (I-PEFiC). I-

PEFiC explains the individual contributions and the

interactions of an agent’s Affordances, Ethics,

Aesthetics, facial Similarity, and Realism to the Use

Intentions and Engagement of the human user. To

date, this framework has an explanatory as well as a

heuristic value because the extracted guidelines are

important for anyone who designs virtual characters.

In a simulation study (Hoorn et al., 2008), we

were capable of formalizing the I-PEFiC framework

and make it the basic mechanism of how agents and

robots build up affect for their human users. In

addition, we designed a special module for affective

decision-making (ADM) that made it possible for

the agent to select actions in favor or against its user,

hence I-PEFiC

ADM

To advance I-PEFiC

ADM

in the area of emotion

regulation, we also looked at other models of affect

(Bosse et al., 2010). Gratch and Marsella (2009)

formalized the theory of Emotion and Adaptation of

Smith and Lazarus (1990) into EMA, to create

agents that cope with negative affect. The emotion-

regulation theory of Gross (2001) inspired Bosse et

al., (2007) to develop CoMERG (the Cognitive

Model for Emotion Regulation based on Gross).

Together, these approaches cover a large part of

appraisal-based emotion theory (Frijda et al.,) and all

three boil down to appraisal models of emotion. We

therefore decided to integrate these three affect

models into a model we called Silicon Coppélia

(Pontier and Siddiqui, 2009; Hoorn et al., 2012).

Figure 1 shows Silicon Coppélia in a graphical

format.

Silicon Coppélia consists of a loop with a

situation as input, and actions as output, leading to a

new situation. This loop consists of three phases: (1)

encoding, (2) comparison, and (3) response.

In the encoding phase, the agent perceives other

agents (whether human or synthetic) in terms of

Ethics (good vs. bad), Affordances (aid vs. obstacle),

Aesthetics (beautiful vs. ugly), and Epistemics

(realistic vs. unrealistic). The agent can be biased in

this perception process, because it is equipped with

desires that have a certain strength for achieving or

preventing pre-defined goal-states (‘get a date’, ‘be

honest’ and ‘connect well’).

In the comparison phase, the agent retrieves

beliefs about actions facilitating or inhibiting the

desired or undesired goal-states to calculate a general

expected utility of each action. Further, agent uses

certain appraisal variables, such as the belief that

someone is responsible for accomplishing goal-states

or not. These variables and the perceived features of

others are appraised for Relevance (relevant or

irrelevant) and Valence to the agent’s goals and

concerns (positive or negative outcome expectancies).

In the response phase of the model, the resulting

appraisals lead to processes of Involvement and

Distance towards the other, and to the emergence of

certain Use Intentions: The agent’s willingness to

employ the other as a tool to achieve its own goals.

Note that both overt (behavioral) and covert

(experiential) responses can be executed in this phase.

Emotions such as hope, joy, and anger are generated

using appraisal variables such as the perceived

likelihood of goal-states. The agent uses an affective

Figure 1: Graphical Representation of Silicon Coppelia.

HowWomenThinkRobotsPerceiveThem-asifRobotswereMen

497

decision-making module to calculate the expected

satisfaction of possible actions. In this module,

affective influences and rational influences are

combined in the decision-making process.

Involvement and Distance represent the affective

influences, whereas Use Intentions and general

expected utility represent the more rational influences.

When the agent selects and performs an action, a new

situation emerges, and the model

starts at the first

phase again.

1.3 Speed-dating as a New Turing-test

In previous research we developed a speed-dating

application as a testbed for cognitive models

(Pontier et al., 2010). In this application, the user

interacted with Tom, a virtual agent on a Website.

We opted for a speed-dating application, because

we expected this domain to be especially useful for

testing emotion models. The emotionally laden

setting of the speed-date simplified asking the user

what Tom would think of them, ethically,

aesthetically, and whether they believed the other

would want to see them again, etc. Further, in a

speed-date there usually is a relatively limited

interaction space; also in our application, where we

made use of multiple choice responses. This was

done to equalize the difference between a human

and our model in the richness of interaction, which

was not our research focus. We wanted the

difference to be based on the success or failure of

our human-like emotion simulations.

We chose to confront female participants with a

male agent, because we expected that the limitations

in richness of behavior in the experiment would be

more easily accepted from a male agent than from a

female one. Previous research suggests that men

usually have more limited forms of emotional

interaction and that women are usually better

equipped to do an emotional assessment of others

(Barret et al., 1998). By means of a questionnaire,

the participants diagnosed the emotional behavior,

and the cognitive structure behind that behavior,

simulated by our model, or performed by a

“puppeteer” controlling Tom.

A pilot study (Pontier et al., 2010) showed that

users recognized at least certain forms of human

affective behavior in Tom. Via a questionnaire, users

diagnosed for us how Tom perceived them and

whether they recognized human-like affective

mechanisms in Tom. Although Tom did not

explicitly talk about it, the participants recognized

human-like perception mechanisms in Tom’s

behavior. This finding was a first indication that our

software had a humanoid way of assessing humans,

not merely other software agents.

These results made us conduct a follow-up

‘Wizard of Oz’ (Landauer, 1987) experiment with

54 participants. In this experiment we compared the

performance of Tom equipped with Silicon Coppélia

to the performance of a human controlling Tom as a

puppeteer. This experiment may count as an

advanced version of a Turing Test (Turing, 1950).

In a Turing Test, however, participants are

routinely asked whether they think the interaction

partner is a human or a robot. In this experiment,

however, we did not ask them so directly. After all,

because of the limited interaction possibilities of a

computer interface, the behavior of Tom may not

seem very human-like. Therefore, all participants

would probably have thought Tom was a robot, and

not a human, making it impossible to measure any

differences. Therefore, we introduced the speed-

dating partner as a robot to see whether humans

would recognize human affective structures equally

well in the software and in the puppeteer condition.

Further, when testing the effect of a virtual

interaction partner on humans, participants are usually

asked how they experience the character. In this

experiment, however, we asked people how they

thought the character perceived them. Thus, the

participants served as a diagnostic instrument to assess

the emotional behavior of Tom, and to detect for us the

cognitive structure behind that behavior. This way, we

could check the differences between our model and a

human in producing emotional behavior, and the

cognitive structure responsible for that behavior.

We hypothesized that we would not find any

differences between the behavior of Tom controlled

by our model and that of Tom controlled by a

human, indicating the success of Silicon Coppélia as

a humanoid model of affect generation and

regulation. This would also indicate the aptness of

the theories the model is based on. Because Silicon

Coppélia is computational, this would also be very

interesting for designing applications in which

humans interact with computer agents or robots.

2 METHOD

2.1 Participants

A total of 54 Dutch female heterosexual students

ranging from 18-26 years of age (M=20.07,

SD=1.88) volunteered for course credits or money (5

Euros). Participants were asked to rate their

experience in dating and computer-mediated

communication on a scale from 0 to 5. Participants

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

498

communicated frequently via a computer (M = 4.02,

SD = 1.00) but appeared to have little experience in

online dating (M = .33, SD = .80).

2.2 Materials:

Speed-dating Application

We designed a speed-date application in which users

could interact with a virtual agent, named Tom, to

get acquainted and make an appointment. The dating

partner was represented by Tom, an avatar created in

Haptek’s PeoplePutty software.

Tom is capable of simulating five emotions:

hope, fear, joy, distress, and anger, which were

expressed through the face of the avatar with either a

low or a high intensity. This depended on little or

much relevance of user choices to Tom’s goals and

concerns. Like this, we created 32 (2

) different

emotional states in PeoplePutty; one for each

possible combination of two levels of intensity of

the five simulated emotions.

We created a Web page for the application (see

Figure 2), in which the virtual agent was embedded

as a Haptek player. We used JavaScript in

combination with scripting commands provided by

the Haptek software, to control the Haptek player

within the Web browser. In the middle of the Web

site, the affective conversational agent was shown,

communicating messages through a voice

synthesizer (e.g., “Do you have many hobbies?”)

and additionally shown as text right above the

avatar. Figure 2 shows that the avatar looks annoyed

in response to the user’s reply “Well, that’s none of

your business”.

Figure 2: The speed-dating application.

During the speed-date, partners could converse

about seven topics: (1) Family, (2) Sports, (3)

Appearance, (4) Hobbies, (5) Music, (6) Food, and

(7) Relationships. For each topic, the dating partners

went through an interaction tree with responses that

they could select from a dropdown box. To give an

idea of what the interaction trees look like, we

inserted the tree for Relationships in the Appendix.

When the ‘start speed-date’ button above the text

area was pressed, Tom introduced himself and

started by asking the user a question. The user

selected an answer from the dropdown box below

Tom. Then Tom responded and so on until the

interaction-tree was traversed. When a topic was

done, the user could select a new topic or let Tom

select one. When all topics were completed, the

message “the speed-dating session is over” was

displayed and the user was asked to fill out the

questionnaire.

In the speed-dating application, Tom perceived

the user according to Silicon Coppélia (Hoorn et al.,

2012). Tom had beliefs that features of the user

influenced certain goal-states in the world. For our

speed-date setting, the possible goal-states were ‘get

a date’, ‘be honest’, and ‘connecting well’ on each

of the conversation topics. Tom had beliefs about the

facilitation of these goal-states by each possible

response. Further, Tom attached a general level of

positivity and negativity to each response.

During the speed-date, Tom updated its

perception of the user based on her responses during

the speed-date, as described in (Pontier et al., 2010).

The assessed Ethics, Aesthetics, Realism, and

Affordances of the user led, while matching these

aspects with the goals of Tom, to Involvement and

Distance towards the human user and a general

expected utility of each action. Each time, Tom

selected its response from a number of options. The

expected satisfaction of each possible response was

calculated based on the Involvement and Distance

towards the user and the general expected utility of

the response, using the following formula:

ExpectedSatisfaction

(Action)

* GEU

(Action)

pos

* (1 - abs(positivity – bias

* Involvement)) +

neg

* (1 - abs(negativity – bias

* Distance))

Tom searched for an action with the level of

positivity that came closest to the level of

Involvement, with the level of negativity closest to

the level of Distance, and with the highest expected

utility (GEU). Tom could be biased to favor positive

or negative responses to another agent.

During the speed-date, Tom simulated a series of

emotions, based on the responses given by the user.

Hope and fear were calculated each time the user

gave an answer. Hope and fear of Tom were based

on the perceived likelihood that he would get a

HowWomenThinkRobotsPerceiveThem-asifRobotswereMen

499

follow-up date. The joy and distress of Tom were

based on achieving desired or undesired goal-states

or not. The anger of Tom was calculated using the

assumed responsibility of the human user for the

success of the speed-date.

All five emotions implemented into the system

(i.e., hope, fear, joy, distress, and anger) were

simulated in parallel. If the level of emotion was

below a set boundary, a low intensity of the emotion

was facially expressed by Tom. If the level of

emotion was greater or equal than the boundary, a

high intensity of the emotion was expressed by Tom.

2.3 Design

The participants were randomly assigned to two

experimental conditions. In the first condition, Tom

was controlled by Silicon Coppélia, whereas in the

second condition Tom was controlled by a human

trained to handle him (Wizard of Oz condition,

WOz). All participants assumed they were

interacting with a robotic partner; also in the WOz

condition. To have some control over the

idiosyncrasies of a single human controller, the

WOz condition consisted of two identical sub-

conditions with a different human puppeteer in each.

Thus, we had three conditions: (1) Tom was

controlled by Silicon Coppélia (n=27), (2) Human 1

controlled Tom (n=22), (3) Human 2 controlled Tom

(n=5). Taken together, 27 participants interacted

with an agent controlled by a human, and 27

participants interacted with an agent controlled by

our software. This way, the behavior simulated by

our model could be compared to behavior of the

human puppeteers. In other words, this was an

advanced kind of Turing Test where we compared

the cognitive-affective structure between conditions.

In a traditional Turing Test, participants do not know

whether they interact with a computer or not

whereas in our set-up participants were told they

were interacting with a robot to avoid rejection of

the dating partner on the basis of limited interaction

possibilities.

2.4 Procedure

Participants were asked to take place behind a

computer. They were instructed to do a speed-date

session with an avatar. In the WOz, the human

controlling the avatar was behind a wall, and thus

invisible for the participants. After finishing the

speed-dating session of about 10 minutes, the

participants were asked to complete a questionnaire

on the computer. After the experiment, participants

in the WOz were debriefed that they were dating an

avatar controlled by a human.

2.5 Measures

The questionnaire consisted of 97 Likert-type items

with 0-5 rating scales, measuring agreement to

statements. Together there were 15 scales. We

designed five emotion scales for Joy, Anger, Hope,

Fear, and Sadness, based on (Wallbot & Scherer,

1989). We also designed a scale for Situation

Selection, with items such as ‘Tom kept on talking

about the same thing’ and ‘Tom changed the

subject’, and a scale for Affective Decision-Making,

with items such as ‘Tom followed his intuition’ and

‘Tom made rational choices’. For all eight

parameters that were present in the I-PEFiC model

(Ethics, Affordances, Similarity, Relevance,

Valence, Involvement, Distance, Use Intentions), the

questions from previous questionnaires (e.g., Van

Vugt, Hoorn & Konijn, 2009) were adjusted and

reused. However, because of the different

application domain (i.e. speed dating), and because

the questions were now about assessing how Tom

perceived the participant, and not about how the

participant perceived Tom, we found it important to

check the consistency of these scales again.

A scale analysis was performed, in which items

were removed until an optimal Cronbach’s alpha

was found and a minimum scale length of three

items was achieved. If removing an item only

increased Cronbach’s alpha very little, the item was

maintained. After scale analysis, a factor analysis

was performed, to check divergent validity. After

additional items were removed, again a scale

analysis was performed (Appendix). All alphas,

except those for Ethics and Similarity, were between

.74 and .95. The scale for Similarity had an alpha of

.66. Previous studies showed that the present Ethics

scale was consistently reliable, and an important

theoretical factor. Therefore, we decided to maintain

the Ethics scale despite its feeble measurement

quality.

2.6 Statistical Analyses

We performed a multivariate analysis of variance

(MANOVA) on the grand mean scores to scales, to

test whether the participants perceived a difference

in Agent-type (software vs. human controlled). We

performed paired t-tests for related groups of

variables.

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

500

3 RESULTS

3.1 Emotions

To analyze the differences in perceived emotions in

the three agent types, we performed a 3x5

multivariate analysis of variance (MANOVA) of the

between-factor Agent-type (3: Silicon Coppélia,

Human1, Human2) and the within-factor of Emotion

(5: Joy, Sadness, Hope, Fear, Anger) on the grand

mean scores to statements. The main effect of

Agent-type on the grand mean scores to emotion

scales was not significant (F

(2, 51)

= 1.68, p < .196),

whereas the main effect of the Emotion factor was

significant (Pillai’s Trace = .64, F

(4, 48)

= 21.59, p <

.001, 

= .64). The interaction between Agent-type

and Emotions was not significant (Pillai’s Trace =

.22, F

(8, 98)

= 1.545, p < .152). More detailed results

can be found in the Appendix.

Because the main effect of Agent-type to

Emotion scales was not significant, this might mean

that there was no effect of emotion at all within a

condition. To check whether emotional behavior was

diagnosed at all by the participants, we performed a

one-sample t-test with 0 as the test value, equalling

no emotions diagnosed. Results showed that all

emotion scales differed significantly from 0. The

smallest t-value was found for Anger (t

(2, 51)

= 8.777,

p < .001).

In addition, the significant main effect of the

Emotion factor suggested that there were systematic

differences in diagnosing emotions in Tom, which

we analyzed by paired samples t-tests for all pairs of

emotions. Out of the 10 thereby originated pairs, 6

pairs differed significantly. The 4 pairs that did not

differ significantly were Joy and Hope (p < .444),

Fear and Sadness (p < .054), Fear and Anger (p <

.908), and Sad and Anger (p < .06). Joy (M = 3.05,

SD = 1.03) and Hope (M = 2.96, SD = .82) were

both recognized relatively much in Tom, whereas

Fear (M=1.04, SD=.80), Sad (M=.84, SD=.66) and

Anger (M=1.02, SD=.86) were recognized little in

Tom.

In other words, the t-tests showed that emotions

were recognized in all conditions, and the

MANOVA showed that participants saw equal

emotions in humans and robots alike.

3.2 Perceptions

To analyze the differences in perceived perceptions

in the three agent-types, we performed a 3x8

MANOVA of the between-factor Agent-type (3:

Silicon Coppélia, Human1, Human2) and the within-

factor of Perception (8: Ethics, Affordances,

Relevance, Valence, Similarity, Involvement,

Distance, Use Intentions) on the grand mean scores

to statements. The main effect of Agent-type on the

perception scale scores was not significant (F < 1),

whereas the main effect of the Perception factor was

significant (Pillai’s Trace = .87, F

(7, 43)

= 39.63, p <

.001, 

= .87). The interaction between Agent-type

and Perception was not significant (Pillai’s Trace =

.18, F

(14, 88)

= .635, p < .828). More detailed results

can be found in the Appendix.

Because the main effect of Agent-type to

Perception scales was not significant, this might

mean that there was no effect of perception at all

within a condition. To check whether the

perceptions of Tom were diagnosed at all by the

participants, we performed a one-sample t-test with

0 as the test value, equalling no perceptions

diagnosed. Results showed that all perception scales

differed significantly from 0. The smallest t-value

was found for Distance (t

(2, 51)

= 15.865, p < .001).

In addition, the significant main effect of the

Perception factor suggested that there were

systematic differences in diagnosing perceptions in

Tom, which we analyzed by paired samples t-tests

for all pairs of perceptions. Out of the 28 thereby

originated pairs, 23 pairs differed significantly. The

pair that differed the most was Ethics and Distance

(t(51) = 13.59, p < .001).

Tom’s perceptions of Ethics (M = 3.86, SD =

.68) and Affordances (M = 3.78, SD = .81) in the

participant were rated the highest. His perceptions of

feeling distant towards the participant (M = 1.77, SD

= .93) were rated the lowest.

In other words, the t-tests showed that

perceptions were recognized in all conditions, and

the MANOVA showed that participants saw equal

perceptions in humans and robots alike.

3.3 Decision-making Behavior

To analyze the differences in perceived decision-

making behavior in the three agent-types, we

performed a 3x2 MANOVA of the between-factor

Agent-type (3: Silicon Coppélia, Human1, Human2)

and the within-factor of Decision-making behavior

(2: Affective decision making, Situation selection)

on the grand mean scores to statements. The main

effect of Agent-type was not significant (F < 1),

whereas the main effect of Decision-making

behavior was small but significant (Pillai’s Trace =

.088, F

(1, 51)

= 4.892, p < .031, 

= .088). The

interaction between Agent-type and Decision-

making behavior was not significant (Pillai’s Trace

HowWomenThinkRobotsPerceiveThem-asifRobotswereMen

501

= .04, F

(2, 51)

= .1.128, p < .332). More detailed

results can be found in the Appendix.

Because the main effect of Agent-type to

Decision-making behavior scales was not

significant, this might mean that there was no effect

of Decision-making behavior at all within a

condition. To check whether decision-making

behavior was diagnosed at all by the participants, we

performed a one-sample t-test with 0 as the test

value, equalling no decision-making behavior

diagnosed. Results showed that both Situation

selection (t

(2, 51)

= 14.562, p < .001) and Affective

decision-making (t

(2, 51)

= 15.518, p < .001) both

differed significantly from 0.

In addition, the significant main effect of the

Perception factor on Agreement suggested that there

were systematic differences in diagnosing

perceptions in Tom, which we analyzed by paired

samples t-test for affective decision-making (M =

2.24, SD = 1.07) and situation selection (M = 1.91,

SD = 1.32). The pair differed significantly (t(53) =

1.776, p < .081).

In other words, the t-tests showed that decision-

making behavior was recognized in all conditions, and

the MANOVA showed that participants saw equal

decision-making behavior in humans and robots alike.

4 DISCUSSION

4.1 Conclusions

In this paper, we equipped a virtual agent with

Silicon Coppélia (Hoorn et al., 2012), a cognitive

model of perception, affection, and affective

decision-making. As an advanced, implicit version

of a Turing Test, we let participants perform a

speed-dating session with Tom, and asked them how

they thought Tom perceived them during the speed-

date. What the participants did not know, was that in

one condition, a human was controlling Tom,

whereas in the other condition, Tom was equipped

with Silicon Coppélia.

A novel element in this experiment was that

participants were asked to imagine how an agent

perceived them. To our knowledge there does not

exist previous research in which participants were

asked to assess the perceptions of an artificial other.

It is a nice finding, that the scales of I-PEFiC (Van

Vugt et al., 2009), which were originally used to ask

how participants perceived an interactive agent,

could be used quite well to ask participants how they

thought Tom perceived them.

The results showed that in this enriched and

elaborated version of the classic Turing Test,

participants did not detect differences between the

two versions of Tom. Not that the variables

measured by the questionnaire did not have any

effect; the effects just did not differ. Thus, within the

boundaries of limited interaction possibilities, the

participants felt that human and software perceived

their moral fiber in the same way, deemed their

relevance the same, and so on. The participants felt

that human and software were equally eager to meet

them again, and exhibited equal ways to select a

situation and to make affective decisions. Also, the

emotions the participants perceived in Tom during

the speed-date session did not differ between

conditions. Emotion effects could be observed by

the participants, and these effects were similar for a

human controlled avatar and software agent alike.

This is good for the engineer who wants to use these

models for application development, such as the

design of virtual agents or robots. After all, on all

kinds of facets, participants may not experience any

difference between the expression of human

behavior and behavior generated by our model.

4.2 Applications

Our findings can be of great use in many

applications, such as (serious) digital games, virtual

stories, tutor and advice systems, or coach and

therapist systems. For example, Silicon Coppélia

could be used to improve the emotional intelligence

of a ‘virtual crook’ that could be used for police

studies to practice situations in which the police

officers should work on the emotions of the crook,

for example questioning techniques (Hochschild,

1983). Another possible use of models of human

processes is in software and/or hardware that

interacts with a human and tries to understand this

human’s states and processes and responds in an

intelligent manner. Many ambient intelligence

systems (e.g., Aarts et al., 2001) include devices that

monitor elderly persons. In settings where humans

interact intensively with these systems, such as

cuddle bots for dementia patients (e.g., Nakajima et

al, 2001), the system can combine the data gathered

from these devices with Silicon Coppélia to maintain

a model of the emotional state of the user. This can

enable the system to adapt the type of interaction to

the user’s needs.

Silicon Coppélia can also be used to improve

self-help therapy. Adding the moral reasoning

system will be very important for that matter.

Humans with psychological disorders can be

supported through applications available on the

Internet and virtual communities of persons with

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

502

similar problems.

New communication technologies have led to an

impressive increase of self-help programs that are

delivered through the Internet (e.g., Spek et al.,

2007). Several studies concluded that self-help

therapies can be more efficient in reducing mental

health problems, and less expensive than traditional

therapy (e.g., Andrews et al., 2001; Bijl and Ravelli,

2000; Cuijpers, 1997; Spek et al., 2007).

Web-based self-help therapy can be a solution

for people who would otherwise not seek help,

wishing to avoid the stigma of psychiatric referral or

to protect their privacy (Williams, 2001). The

majority of persons with a mental disorder in the

general population do not receive any professional

mental health services (an estimated 65%) (Andrews

et al., 2001; Bijl and Ravelli, 2000). In many

occupations, such as the police force, the fire service

and farming, there is much stigma attached to

receiving psychological treatment, and the

anonymity of Web-based self-help therapy would

help to overcome this. Also many other people feel a

barrier to seek help for their problems through

regular health-care systems; e.g., in a study by Spek

et al. (2007) about internet-based cognitive

behavioral therapy for sub-threshold depression for

people over 50 years old, many participants reported

not seeking help through regular health-care systems

because they were very concerned about being

stigmatized. Patients may be attracted to the idea of

working on their own to deal with their problems,

thereby avoiding the potential embarrassment of

formal psychotherapy (Williams, 2001).

Further, self-help therapy is particularly suited to

remote and rural areas, where ready access to a face-

to-face therapist cannot be economically justified.

Self-help therapy may also be useful in unusual

environments such as oilrigs and prisons, where

face-to-face therapy is not normally available. Self-

help therapy can also be offered to patients while

they are on a waiting list, with the option to receive

face-to-face therapy later, if required (Peck, 2007)

Self-help therapy may be even more successful

when the interface is enhanced or replaced by a

robot therapist that has Silicon Coppélia installed.

The anonymity of robot-supported self-help therapy

could overcome potential embarrassment of

undergoing formal treatment. When regular therapy

puts up too high a threshold, a robot therapist is less

threatening, what the patient reveals is

inconsequential, the patient is in control, and all in

all, interaction with the virtual therapist has a “dear

diary” effect. As if you were speed-dating with a

real partner.

4.3 Future Research

In future research, we will test an extended version

of the current model, using robots in the healthcare

domain. So-called Caredroids will play a chess game

with the patient as a form of daytime activity. Based

on whether the agent reaches its goals (winning and

losing when the agent has ambitions to win or lose),

the likelihood of these goals, and the expectedness

of the move of the user and the outcome of a game,

the emotions joy, distress, hope, fear and surprise are

simulated and shown by the agent by means of

bodily expressions. The Caredroid will be able to

trade rational choices to win the game for affective

choices to let the human opponent win if she is nice

to him.

Additionally, we will integrate Silicon Coppélia

with a moral reasoning system that can solve

medical ethical dilemmas (Pontier and Hoorn,

2012). In this system, actions are evaluated against a

number of moral principles to point out ethical

dilemmas in employing robot care.

In entertainment settings, we often like

characters that are naughty; the good guys often are

quite boring (Konijn and Hoorn, 2005). In Silicon

Coppélia (Hoorn et al., 2012), this could be

implemented by updating the affective decision

making module. Morality would be added to the

other influences that determine the Expected

Satisfaction of an action in the decision making

process. By doing so, human affective decision-

making behavior could be further explored. Some

inital steps in doing this were taken in (Pontier,

Widdershoven and Hoorn, 2012).

ACKNOWLEDGEMENTS

This study is part of the SELEMCA project within

CRISP (grant number: NWO 646.000.003). We

would like to thank Marco Otte for his help in

conducting the experiment. We are grateful to Elly

Konijn for valuable discussions and advice.

REFERENCES

Aarts, E., Harwig, R., and Schuurmans, M., 2001.

Ambient Intelligence. In The Invisible Future: The

Seamless Integration of Technology into Everyday

Life, McGraw-Hill, 2001.

Andrews, G., Henderson, S., & Hall, W., 2001.

Prevalence, comorbidity, disability and service

utilisation. Overview of the Australian National

HowWomenThinkRobotsPerceiveThem-asifRobotswereMen

503

Mental Health Survey. In The British Journal of

Psychiatry, Vol. 178, No. 2, pp. 145-153.

Barrett, L. F., Robin, L., Pietromonaco, P. R., Eyssell, K.

M., 1998. Are Women The "More Emotional Sex?"

Evidence from Emotional Experiences in Social

Context. In Cognition and Emotion, 12, pp. 555-578.

Bijl, R. V., & Ravelli, A., 2000. Psychiatric morbidity,

service use, and need for care in the general

population: results of The Netherlands Mental Health

Survey and Incidence Study. In American Journal of

Public Health, Vol. 90, Iss. 4, pp. 602-607.

Bosse, T., Gratch, J., Hoorn, J. F., Pontier, M. A.,

Siddiqui, G. F., 2010. Comparing Three

Computational Models of Affect. In Proceedings of

the 8th International Conference on Practical

Applications of Agents and Multi-Agent Systems,

PAAMS, pp. 175-184.

Bosse, T., Pontier, M. A., Treur, J., 2007. A Dynamical

System Modelling Approach to Gross´ Model of

Emotion Regulation. In Proceedings of the 8th

International Conference on Cognitive Modeling,

ICCM'07, pp. 187-192, Taylor and Francis.

Brave, S., Nass, C., & Hutchinson, K., 2005. Computers

that care: investigating the effects of orientation of

emotion exhibited by an embodied computer agent.

International Journal of Human-Computer Studies,

62(2), pp. 161–178.

Cuijpers, P., 1997. Bibliotherapy in unipolar depression, a

meta-analysis. In Journal of Behavior Therapy &

Experimental Psychiatry, Vol. 28, No. 2, pp. 139-147

Gross, J. J., 2001. Emotion Regulation in Adulthood:

Timing is Everything. In Current Directions in

Psychological Science, 10(6), pp. 214-219.

Hochschild, A. R., 1983. The managed heart. Berkeley:

University of California Press.

Hoorn, J. F., Pontier, M. A., Siddiqui, G. F., 2008. When

the user is instrumental to robot goals. First try: Agent

uses agent. In Proceedings of International

Conference on Web Intelligence and Intelligent Agent

Technology, WI-IAT ‘08, IEEE/WIC/ACM, Sydney

AU, 296-301.

Hoorn, J. F., Pontier, M. A., & Siddiqui, G. F., 2012.

Coppélius' Concoction: Similarity and

Complementarity Among Three Affect-related Agent

Models. Cognitive Systems Research, 2012, pp. 33-49.

Konijn, E. A., & Hoorn, J. F., 2005. Some like it bad.

Testing a model for perceiving and experiencing

fictional characters. Media Psychology, 7(2), 107-144.

Konijn, E. A., & Van Vugt, H. C., 2008. Emotions in

mediated interpersonal communication. Toward

modeling emotion in virtual agents. In Konijn, E.A.,

Utz, S., Tanis, M., & Barnes, S. B. (Eds.). Mediated

Interpersonal Communication. New York: Taylor &

Francis group/ Routledge.

Landauer, T. K., 1987. Psychology as a mother of

invention. Proceedings of the SIGCHI/GI conference

on Human factors in computing systems and graphics

interface, pp. 333-335.

Marsella, M., Gratch, J., 2009. EMA: A Process Model of

Appraisal Dynamics. In Journal of Cognitive Systems

Research, Vol. 10, No. 1, pp. 70-90.

Nakajima, K., Nakamura, K., Yonemitsu, S., Oikawa, D.,

Ito, A., Higashi, Y., Fujimoto, T., Nambu, A., Tamura,

T., 2001. Animal-shaped toys as therapeutic tools for

patients with severe dementia. Engineering in

Medicine and Biology Society. In Proceedings of the

23rd IEEE Conference, vol. 4, 3796-3798.

Peck, D., 2007. Computer-guided cognitive–behavioral

therapy for anxiety states. Emerging areas in Anxiety,

Vol. 6, No. 4, pp. 166-169.

Picard, R. W., 1997. Affective Computing. Cambridge,

MA: MIT Press.

Pontier, M. A., & Hoorn, J. F., 2012. Toward machines

that behave ethically better than humans do. In

Proceedings of CogSci'12, pp. 2198-2203.

Pontier, M. A., Siddiqui, G. F., 2009. Silicon Coppélia:

Integrating Three Affect-Related Models for

Establishing Richer Agent Interaction. In WI-IAT’09,

International Conference on Intelligent Agent

Technology, vol. 2, pp. 279—284.

Pontier, M. A., Siddiqui, G. F., & Hoorn, J., 2010. Speed-

dating with an Affective Virtual Agent - Developing a

Testbed for Emotion Models In: J. Allbeck et al.

(Eds.), In IVA'10, Lecture Notes in Computer Science,

Vol. 6356, pp. 91-103.

Pontier, M. A., Widdershoven, G. A. M., and Hoorn, J. F.,

2012. Moral Coppélia - Combining Ratio with Affect

in Ethical Reasoning. In: Proceedings of the 13th

Ibero-American Conference on Artificial Intelligence,

IBERAMIA’12, in press.

Smith, C. A., & Lazarus, R. S., 1990. Emotion and

adaptation. In L. A. Pervin (Ed.), Handbook of

personality: Theory & research, (pp. 609-637). NY:

Guilford Press.

Spek, V., Cuijpers, P., Nyklíĉek, I., Riper, H., Keyzer, J.,

& Pop, V., 2007. Internet-based cognitive behavior

therapy for emotion and anxiety disorders: a meta-

analysis. In Psychological Medicine, Vol. 37, pp. 1-10

Turing, A. M., 1950. Computing Machinery and

Intelligence. Mind 54(236), pp. 433-460.

Van Vugt, H. C., Hoorn, J. F., & Konijn, E. A., 2009.

Interactive engagement with embodied agents: An

empirically validated framework. In Computer

Animation and Virtual Worlds, Vol. 20, pp. 195-204

Wallbot, H. G., & Scherer, K. R., 1989. Assessing

Emotion by Questionnaire. Emotion Theory, Research

and Experience, Vol. 4, pp. 55-82.

Williams, C., 2001. Use of Written Cognitive-Behavioral

Therapy Self-Help Materials to treat depression. In

Advances in Psychiatric Treatment, Vol. 7, pp. 233–240.

APPENDIX.

http://camera-vu.nl/matthijs/SpeeddateAppendix.pdf

ICAART2013-InternationalConferenceonAgentsandArtificialIntelligence

504