Agents with Faces: The Effect of Personi?cation

TAG:  poker game 

Published Time: -

Filetype: pdf

Filesize: 64197

quantitative analysis. The primary focus of this research is
on implementation of a prototype rather than user testing. However, other research with quantitative analysis indi- cates that adding a face to an interface does not necessarily
result in better human-computer interactions. Walker et al.
[9] report that having a face is engaging and takes more
effort and attention from the user. A face with more expres-
sion leads to greater engagement, but is not always prefera-
ble. Takeuchi et al. report that users respond differently to
systems having a face than to those without. They suggest
that a face in an interface takes more effort from the user
because people try to interpret the human images [8]. The research reported in this paper includes both imple- mentation and quantitative analysis of subjects’ impression
about a personi?ed interface. This paper argues that
employing a face as the representation of an agent is engag-
ing and makes a user pay attention. It is possible to make
advantages of personi?cation outweigh the drawbacks by
using faces in appropriate application domains, where con-
veying the agent's behavior to the user is important for mak-
ing a user comfortable with the system, or in an
entertainment domain where engagement is crucial and
users don't try to minimize efforts. This paper reports the
results of an investigation on people's impressions of faces
in an interactive entertainment environment. 2. Experiment We built an application to test agent-based interfaces, namely a poker game, and experiments were performed on
the World Wide Web so as to have as many versatile sub-
jects as possible. The results were analyzed quantitatively. 2.1 Purpose The purpose of the experiment is to understand 1) the effect of having a face and facial expressions in an interface,
such as required attention, engagement, and distraction, 2)
what kind of facial features (gender, humanity, realism)
make the agent look intelligent, likable, and comfortable to
work with, 3) whether people’s impression of an agent is Agents with Faces: The Effect of Personi?cation Tomoko Koda / Pattie Maes MIT Media Laboratory 20 Ames Street Cambridge, MA 02139 USA tomoko/pattie@media.mit.edu Abstract It is still an open question whether software agents should be personi?ed in the interface. In order to study the
effects of faces and facial expressions in the interface, a
series of experiments was conducted to compare subjects’
responses to and evaluation of different faces and facial
expressions. The experimental results obtained demonstrate that: 1) personi?ed interfaces help users engage in a task, and are
well suited for an entertainment domain; 2) people’s
impressions of a face in a task are different from ones of the
face in isolation. Perceived intelligence of a face is deter-
mined not by the agent’s appearance but by its competence;
3) there is a dichotomy between user groups which have
opposite opinions about personi?cation. Thus, agent-based
interfaces should be ?exible to support the diversity of
users’ preferences and the nature of tasks. 1. Introduction There is a growing interest in interface agents which assist a user in performing daily computer tasks. Software
agents will make software more active and work autono-
mously without waiting for a user's command. This “agent-
based” interface should be social and active to support
cooperative tasks between a user and an agent [5]. Maes
addresses that the key HCI issue in building a successful
interface agent is to help the user understands what the
agent's capacities are, its limitations, and its way of operat-
ing, etc. [4]. Many software agents are personi?ed in terms of a human or caricature face. “Phil” in Apple's Knowledge
Navigator video [3], Apple's “Guides” project [6], the MIT
Media Lab’s Maxims system [2] used faces to help users
understand the agent's characteristics or behaviors. The
above research hypothesizes that users interact with agents
more easily by being able to predict the agent's characteris-
tics and behaviors via their external traits. Though the above
research reports positive results for a personi?ed interface,
the results are obtained by observation or interviews, not by determined by its representation or by its performance or
both, and 4) whether people’s impression of the faces differs
by their gender or opinion about personi?cation. 2.2 System Design We designed and built a web-based poker game in which computer poker playing agents play against each
other and a user. The task for the user is to play this poker
game against the poker playing agents which have personi-
?ed representations. Figure 1 shows the interface of the
game. The poker playing agents used in the experiments
have seven different graphical representations: Realistic
Human faces, Caricature Human faces, a Caricature Dog’s
face, a line-drawn Smiley face, and the Invisible Man,
which has no face (see Figure 2). They vary in terms of
humanity and realism. Each agent has ten facial expressions
(neutral, pleased, displeased, satis?ed, surprised, excited,
anxious, and very excited). It changes its expressions when
it deals, bets, and wins/loses. For example, a dog’s face
might show an anxious expression while betting, or show an
excited face when bluf?ng. Figure 3 shows the ten expres-
sions of the female caricature face. The emotions were
derived from the Ortony, Clore, and Collins (OCC) model
of emotion types [7]. 2.3 Procedure A typical scenario of an experiment is as follows: A subject accesses the poker web site. The subject reads the consent form and instructions and agrees to partic-
ipate in the experiment. The subject’s name, email address,
and login time are sent to the server and stored in a disk.
The poker java applet is downloaded to the subject’s com-
puter. The subject plays a poker game with one out of ?ve
conditions for 15 rounds, which takes from 10 to 15 minutes
to complete. The subject’s and agents’ hands, actions, and
the ending money state are logged into the server. A ques-
tionnaire appears after playing 15 rounds. The subject
answers and submits the questionnaire, and the answers are
stored in the server. 2.4 Conditions of the Experiments There are ?ve web-based experiments which are ran- domly assigned to a user. The FACE experiment is designed
to study the effect of having a face. The representations are
the Caricature Female face and a white square with the same
size which is called “Invisible Man”. The GENDER experi-
ment studies the difference in subjects’ impressions of char-
acters’ gender. The representations used in this experiment
are the Realistic Male and the Realistic Female face. The
HUMANITY experiment is designed to compare the differ-
ence in subjects’ impressions between a human face and a
non-human face. The representations used are the Carica- ture Male face and the Dog’s face. The REALISM experi- ment uses three faces to compare three different levels of
realism. The representations used are the Realistic Male Figure 1: Web interface of the poker game
(REALISM experiment) Figure 2: The graphical representations of the
poker playing agents Figure 3: Ten facial expressions of the Caricature
Female face face, the Caricature Male face, and the Smiley face. The
EXPRESSIVENESS experiment has three players with the
same facial representation. This experiment was designed to
evaluate the difference in subjects’ impression between
three modes of expressiveness. One agent has Honest
expressions, the other has Deceiving expressions, and the
last one has Stoic expressions. Table 1 shows the facial rep-
resentations used in each experiment. All facial images used
in the experiments have the same size, same resolution, and
the same background color. Except for the EXPRESSIVE-
NESS experiment, all poker playing agents are in Honest
mode. The agents play poker game using exactly the same
poker playing strategy, which the subjects were not made
aware of. Table 1: The facial representations used in each
experiment Three pilot experiments were conducted prior to the ?ve web-based experiments. The purpose of the pilot exper-
iments was to collect subjects’ impressions of each face
based solely upon its visual appearance. Subjects were
shown one of the sets of faces used in the GENDER,
HUMANITY, and REALISM experiments with a neutral
expression. The subjects answered questions about each
face’s intelligence, likability, and engagingness as a oppo-
nent poker player, just by looking at the facial images. 2.5 Questionnaire Questions varied according to the experiment. There are 4 background questions (subjects’s gender, age, com-
puter expertise, and their opinion about personi?cation), and
4 to 8 questions about the subjects’ impressions of the faces
against which they played poker. The questions about their
impressions are answered on a seven point scale (1: strongly
disagree to 7: strongly agree). The common questions across the experiments are per- ceived intelligence of playing poker game, likability, engag-
ingness, and level of comfort for each face. Speci?c
questions are attention required and level of distraction (for
FACE and EXPRESSIVENESS experiments), level of cor-
respondence to the actual intelligence of the player (for
HUMANITY and REALISM experiments), believability of
the facial expressions, and the usefulness of the facial
expressions (for EXPRESSIVENSS experiment). 2.6 Subjects Experiments continued for one and a half months. Sub- jects participated in the experiments from all over the world
using the World Wide Web. More than 1,000 people
accessed the poker game site and 157 of them answered the
questionnaire, for a response rate of 15%. Of these subjects,
78% were male. The age range of the subjects was from 10
to 50 years old. Fifty-seven percent of them were in their
20’s, 26% were teens, and 14% were in their 30’s. Fifty-two
percent of them were advanced computer users, 40% were
intermediate users. When asked about personifying an inter-
face (VOTE), 51% of them supported having a face on the
screen (“AGREE” group), the rest were against having a
face (“DISAGREE” group). 3. Results 3.1 Impressions Based on Appearance The pilot experiments evaluate subjects’ impressions based solely on agents’ appearance, in particular, gender,
humanity, and realism. The result of the comparison for
gender shows that there is no difference between people’s
impression of the Male face and Female face in terms of its
perceived intelligence, likability, and engagingness, as
shown in Table 2. The result of the comparison for humanity
shows that the Human face is perceived as more intelligent
than the Dog’s face based on visual appearances, but less
likable and engaging as a representation for a poker player,
as shown in Table 3. In the experiment to compare different
levels of realism, subjects rated the Realistic face to be more
intelligent, likable, and engaging than the Caricature face
and Smiley, as shown in Table 4. Table 2: The mean value of variables for a Male and
Female face based on appearance * p<.01, ** p<.05, *** p<.10, no mark = not signi?cant Table 3:The mean value of variables for a Human
and Dog’s face based on appearance * p<.01, ** p<.05, *** p<.10, no mark = not signi?cant Player 1 Player 2 Player 3 FACE Caricature Female Invisible man - GENDER Realistic Male Realistic Female - HUMANITY Caricature Male Caricature Dog - REALISM Realistic Male Caricature Male Smiley EXPRES-
SIVENESS Honest Caricature
Female Deceiving Carica-
ture Female Stoic Caricature
Female Variables Male
(n=15) Female
(n=15) t (paired) INT (Intelligence) 4.73 5.13 t(14) = -.88 LIKE (Likability) 4.80 4.73 t(14) =.19 ENG (Engagingness) 5.13 4.80 t(14) =.77 Variables Human
(n=16) Dog
(n=16) t (paired) INT (Intelligence) 3.88 2.63 t(15) = 3.10* LIKE (Likability) 3.81 5.37 t(15) = -3.13* ENG (Engagingness) 3.50 5.56 t(15) = -5.57* Table 4:The mean value of variables for a Smiley,
Caricature, and Realistic face based on
appearance * p<.01, ** p<.05, *** p<.10, no mark = not signi?cant 3.2 Difference for Face and Noface As shown in Table 5, subjects paid attention to the face (ATT = 5.28, where 4.0 is neutral), but were not distracted
by the existence of the face (DIS = 2.84, where 4.0 is neu-
tral). Subjects rated having a face to be signi?cantly more
likable (t(24)=2.76, p<.05), more engaging (t(24)=2.85,
p<.01), and more comfortable to play against (t(24)=3.36,
p<.01), than not having a face. However, both poker playing
agents are rated to be equally intelligent regardless of the
existence of a face. Table 5:The mean value of variables for a
Caricature face(Face) and the Invisible Man
(NoFace) in the game * p<.01, ** p<.05, *** p<.10, no mark = not signi?cant 3.3 Difference for Characters’s Gender As shown in Table 6, there is no main effect of the agent’s gender in any variables. However, there are signi?-
cant interactions between VOTE (people’s opinion about
personi?cation) and perceived intelligence (F(4.73, 1), p
<.05), and slight interactions between VOTE and likabil-
ity(F(3.62, 1), p <.10). The AGREE group (those who are
for personi?cation) rated intelligence and likability for the
Male face higher and the Female face lower. While the DIS-
AGREE group (those who are against personi?cation) rated
the Male and the Female face in the opposite way. No other
2-way or 3-way interactions were found. 3.4 Difference for Degree of Humanity As shown in Table 7, there is no main effect of HUMANITY (Human or Dog’s face) in any variables.
However, there are signi?cant interactions between VOTE and likability (F(5.40, 1), p<.05), comfortableness (F(11.42,
1), p <.01), and level of correspondence to the actual poker
playing skill (F(6.97, 1), p <.05). The AGREE group rated
likability, comfortableness, and correspondence of the
Human face signi?cantly lower while they rated the Dog’s
face signi?cantly higher. The DISAGREE group rated the
Human and the Dog’s face in the opposite way. There are
also interactions between subjects’ gender and likability
(F(9.24, 1), p <.01), engagingness (F(5.94, 1), p <.05), and
comfortableness (F(7.26, 1), p <.05). Table 6: The mean value of variables for a Male and
Female face in the game. * p<.01, ** p<.05, *** p<.10, no mark = not signi?cant 3.5 Difference for Degree of Realism As shown in Table 8, there is no main effect of REAL- ISM in perceived intelligence, likability, and engagingness. Variables Smiley
(n=16) Caricature
(n=16) Realistic
(n=16) F(value, df) INT (Intelligence) 3.53 3.59 4.94 F (6.12, 2)* LIKE (Likability) 3.76 4.06 4.88 F (5.06. 2)** ENG (Engagingness) 3.65 3.82 5.12 F (6.25, 2)* Variables Face
(n=25) NoFace(n
=25) t (paired) INT (Intelligence) 3.64 3.80 t(24) = -.27 LIKE (Likability) 5.08 3.72 t(24) = 2.76** ENG (Engagingness) 4.60 3.32 t(24) = 2.85* COM (Comfortableness) 4.88 3.36 t(24) = 3.06* ATT (Attention to the face) 5.28 - - DIS (Distracted by the face) 2.84 - - Variables Male
(n=37) Female
(n=37) t (paired)
F(value, df) INT (Intelligence) Vote Agree (n=18) Vote Disagree (n=19) 3.35 3.94 2.78 3.83 3.55 4.10 t(36) = -1.33 F(4.73, 1)** LIKE (Likability) Vote Agree (n = 18) Vote Disagree (n = 19) 4.49 5.11 3.89 4.62 4.83 4.42 t(36) = -.55 F(3.62,1)*** ENG (Engagingness) 4.32 4.43 t(36) = -.55 COM (Comfortableness) 4.00 3.92 t(36) =.30 Table 7:The mean value of variables for Human
and Dog’s face in the game Variables Human
(n=23) Dog
(n=23) t (paired)
F(value, df) INT (Intelligence) 3.73 3.82 t(22) = -.18 LIKE (Likability) Vote Agree (n = 11) Vote Disagree (n = 12) Male subject (n = 17) Female subject (n = 6) 4.56 4.72 4.45 4.47 4.83 4.39 4.73 4.09 4.65 3.67 t(22) =.53 F(5.40, 1)** F(9.24, 1)* ENG (Engagingness) Male subject (n = 17) Female subject (n = 6) 3.95 3.88 4.17 4.13 4.29 3.67 t(22) = -1.00 F(5.94, 1)** COM (Comfortableness) Vote Agree (n = 11) Vote Disagree (n = 12) Male subject (n = 17) Female subject (n = 6) 4.08 3.91 4.27 4.29 3.50 4.04 4.55 3.55 4.41 3.00 t(22) =.13 F(11.42, 1)* F(7.26, 1)** COR (Correspondence) Vote Agree (n = 11) Vote Disagree (n = 12) 3.91 3.82 4.00 3.50 4.18 2.82 t(22) = 1.16 F(6.97, 1)** The results show that the Realistic face is slightly more
intelligent, likable, engaging than other faces, though not
signi?cantly so. There is a main effect of REALISM in
comfortableness (F(.05, 2), p <.10) and a signi?cant main
effect in correspondence to the skill level (F(.18, 2), p <.05).
The Realistic face is rated as more comfortable to play
against than the Smiley face. In terms of correspondence to
the actual poker playing skill, subjects rated the Smiley face
as the most appropriate representation, and the caricature
and Realistic faces are less appropriate. No 2-way or 3-way
interactions were found. Table 8: The mean value of variables for Smiley.
Caricature, and Realistic face in the game * p<.01, ** p<.05, *** p<.10, no mark = not signi?cant 3.6 Difference between Levels of Expressiveness As shown in Table 9, there is no main effect of EXPRESSIVENESS in any variables. However, there are
signi?cant interactions between VOTE and perceived intel-
ligence (F(8.91, 2), p <.01) and comfortableness(F(4.47, 2),
p <.05). The AGREE group rated the Honest face as most
intelligent, while the DISAGREE group rated the Stoic face
as the most intelligent and the Honest face as the least. The
AGREE group rated the Stoic face as least comfortable,
while the DISAGREE rated the same face as most comfort-
able. 4 Discussion Subjects rated a poker playing agent with and without a face as equally intelligent. This suggests that attaching a
face did not add any perceived intelligence to the poker
player. However, having a face is considered more likable,
engaging, and comfortable to play against regardless of sub-
jects’ opinion about personi?cation. As described earlier,
Takeuchi’s [8] and Walker’s [9] studies show a similar
result, that having a face is engaging. It is encouraging that
the face did not create a negative effect on the subjects’ lik-
ability of the experience, even to those who have a negative Table 9:The mean value of variables for Honest,
Deceiving, and Stoic face in the game * p<.01, ** p<.05, *** p<.10, no mark = not signi?cant opinion about personi?cation, contrary to what Walker’s
study found [9]. This difference may be caused by the
nature of the tasks and faces used in these studies. Walker
used a woman’s synthetic talking face in a questionnaire
survey, while this study used a female caricature face in a
poker game environment. Subjects’ responses to the Human
and Dog’s face when seen in isolation show reasonable
expectations from the appearance of each face regardless of
their opinion about personi?cation. King’s study [1] shows
a similar result in perceived intelligence, that human forms
are perceived to be more intelligent than other forms. How-
ever, subjects rated the Human and Dog’s faces to be
equally intelligent (or unintelligent) in the poker game. This
means that they did not rate a player’s intelligence based on
its appearance. Their impressions are based on the player’s
actual competence in playing poker. However, their opin-
ions about personi?cation affect their attitude toward each
face. The AGREE group are more “strict” about personi?ca-
tion, since they prefer the Dog’s face to the Human face, felt
more comfortable with the Dog’s face, and considered the
Dog’s face more appropriate for representing the actual
competence of the player in playing poker. On the other
hand, the DISAGREE group are more “generous” about
personi?cation. The subject’s gender works in the same way
as the subjects’ opinion about personi?cation. Males are
more “strict”, while female are more “generous” about per-
soni?cation. However, due to the small number of female
subjects, further study is needed to conclude that users’ gen-
der affects their impressions of agent’s personi?cation. In terms of realism, the Realistic face was rated as more intelligent, engaging, and likable than the less realistic faces
when seen in isolation. Subjects categorized the faces in two
ways--realistic face and not-realistic (Caricature and Smi-
ley) when they evaluate the faces based on appearance. The Variables Smiley
(n=30) Caricat
ure
(n=30) Realist
ic
(n=30) F(value, df)
t (paired) INT (Intelligence) 3.37 3.53 3.70 F (.30, 2) LIKE (Likability) 4.20 4.10 4.70 F (.59, 2) ENG (Engagingness) 4.17 4.13 4.76 F (.58, 2) COM (Comfortableness) Smiley vs. Caricature Caricature vs. Realistic Smiley vs. Realistic 4.07 4.00 4.63 F (.05, 2)*** t(29) =.14 t(29) = -1.60 t(29) = -1.71*** COR (Correspondence) Smiley vs. Caricature Caricature vs. Realistic Smiley vs. Realistic 4.67 3.90 4.03 F (.18, 2)** t(29) = 2.25** t(29)= -.36 t(29) = 2.00** Variables Honest
(n=20) Deceivi
ng
(n=20) Stoic
(n=20) F(value, df) INT (Intelligence) Vote Agree (n = 10) Vote Disagree (n = 10) 3.84 4.20 3.44 3.37 3.10 3.66 3.89 3.30 4.55 F(1.98, 2) F(8.91, 2)* LIKE (Likability) 4.37 4.37 4.68 F(1.28, 2) ENG (Engagingness) 3.95 4.21 4.21 F(.56, 2) COM (Comfortableness) Vote Agree (n = 10) Vote Disagree (n = 10) 3.95 3.90 4.00 3.74 4.10 3.33 3.95 3.60 4.33 F(.14, 2) F(4.47, 2)** BEL (Correspondence) 4.79 4.79 4.63 F(6.44, 2) ATT (Required attention) 4.32 4.47 4.37 F(1.46, 2) DIS (distraction) 3.26 3.21 3.37 F(5.26, 2) USE (usefulness) 4.21 4.42 4.11 F(1.81, 2) level of realism does not affect the overall impression of
faces, unless they are not realistic. This result is similar to
King’s study [1] which indicates that fully articulated
human forms are rated to have higher intelligence and
agency than either the caricatures or the Chernoff faces
(corresponding to Smiley face in this experiment), and cari-
catures and Chernoff faces are rated to have similar intelli-
gence and agency. However, subjects rated the three faces
differently after playing poker in this study. They judged the
level of perceived intelligence of a face based not on its
appearance but instead on its competence. Subjects rated the
three faces to be equally intelligent, and they thought that
the Smiley face (less realistic) face represented the true level
of competence more appropriately. It is interesting that people’s opinion about personi?ca- tion affects the way they feel about expressiveness. The
AGREE group rated expressive faces (Honest and Deceiv-
ing) to be more comfortable, and the Honest face to be most
intelligent. The DISAGREE group rated the stoic face as
most intelligent and comfortable. This means the DIS-
AGREE group is more strict about simulating a real poker
situation, where players are not expected to show honest
expressions. 5 Conclusion The ?rst ?nding is that having a face is considered more likable, engaging, and comfortable to see in a poker
game environment. It is encouraging that people are favor-
able to having a face in an interface regardless of their opin-
ion about personi?cation. People were not distracted by the
presence of a face or facial expressions. Moreover, people
tried to interpret faces and facial expressions, which makes
the users pay attention to the face and engage in the task. It
is clear that faces are useful for entertainment purposes,
since engagement is essential for games and people don’t
care about taking more effort for entertainment. Hence it
may also be useful for applications which require an
engaged user for success, such as education and training. The second ?nding is that people’s impressions of a face are different when they see a face in isolation versus
when they interact with a face within a task. People evaluate
a face not based on appearance but its competence or perfor-
mance. For example, people rated perceived intelligence,
likability, and engagingness of the Human face and the
Dog’s face, or Smiley, the Caricature, and the Realistic face
differently when they rated the faces based on their appear-
ance. While there were no differences in evaluations when
they rated the same faces in the poker game. Most psycho-
logical HCI studies use static facial images separately from
applications. Considering what we understand from this
study, we have to evaluate effects of personi?cation within a
context,--i.e. the context of software agent applications. Of
course not all software agents require personi?ed interfaces.
The goal of HCI work should be to understand when a per- soni?ed interface is appropriate. The third ?nding is that there is a dichotomy between user groups which have opposite opinions about personi?-
cation. Differences in facial features--such as character’s
gender, humanity, expressiveness--cause opposite evalua-
tions by these two subject groups. For example, the Dog’s
face is preferred by those who are in favor of personi?ca-
tion, while the Human face is preferred by those who are
against. Another example is that those who are in favor of
personi?cation attributed more intelligence to the expressive
face, while those who are against thought the stoic face had
higher level of intelligence. Some of the experiments per-
formed indicate that there is also a potential difference in
evaluation of a human face and a non-human face between
the subjects’ gender. Subjects’ computer expertise, age
range were not taken into consideration in this study. We
need to consider the target users when designing a personi-
?ed interface. The future personi?ed interface should be
?exible so that it can provide options to choose a preferred
face for each user. Acknowledgments This research was supported by Apple Computer and British Telecom. References 1. King, J. et al. “The Representation of Agents: Anthropomor-
phism, Agency, and Intelligence.” in Proc. of CHI ‘96. ACM
Press, 1996. 289-290. 2. Lashkari, Y. et al. “Collaborative Interface Agents.” In Proc. of
the National Conference on Arti?cial Intelligence. Cambridge,
MA.: MIT Press, 1994 3. Laurel, B. “Interface Agents: Metaphors with Character.” In
The Art of Human-computer Interface Design, edited by
Brenda Laurel. Reading, MA: Addison-Wesley, 1990. 355-
365. 4. Maes, P. “Agents That Reduce Work and Information Over-
load”. In Communications of the ACM. Vol. 37, No. 7 (July,
1994). 31-40,146. 5. Negroponte, N. Being Digital. 101-102. The Alfred A.
Knorpf, Inc. 1995. 6. Oren, T. et al. “Guides: Characterizing the Interface.” In The
Art of Human-computer Interface Design, edited by Brenda
Laurel. Reading, MA: Addison-Wesley, 1990. 367-381. 7. Ortony, A. et al. “The Cognitive Structure of Emotions.” Cam-
bridge, MA: Cambridge University Press. 1988 8. Takeuchi, A. et al. “Situated Facial Displays: Towards Social
Interaction,” in Proc. of CHI ‘94, ACM Press, 1994. 450-454. 9. Walker, J. et al. “Using a Human Face in an Interface,” in Proc.
of CHI ‘94, ACM Press, 1994. 85-91.